<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[CData Software - Medium]]></title>
        <description><![CDATA[Real-time connectivity, data replication &amp; managed AI integration for 350+ sources. Secure, governed access for analytics, ops &amp; agents — https://www.cdata.com/ - Medium]]></description>
        <link>https://medium.com/cdata-software?source=rss----47213ef81ae5---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>CData Software - Medium</title>
            <link>https://medium.com/cdata-software?source=rss----47213ef81ae5---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 15:40:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/cdata-software" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Agent Memory Architecture: Short- & Long-Term Governed Access]]></title>
            <link>https://medium.com/cdata-software/agent-memory-architecture-short-long-term-governed-access-a7b14833a0d6?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/a7b14833a0d6</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 17:23:25 GMT</pubDate>
            <atom:updated>2026-11-05T17:23:25.776Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*McoA5FYnAroJv9S2" /></figure><p>Working memory and episodic logging are solved problems: the context window holds the current task, a checkpointer saves the conversation thread, and logs record what the agent did in prior sessions. When the agent reaches into a CRM, a data warehouse, or an HR system for facts it needs, that retrieval rarely checks whether the requesting user has permission to see what comes back. The issue is that the agent can return a record the requesting user isn’t cleared to see in the source system.</p><h3>What is agent memory architecture?</h3><blockquote><strong><em>Quick definition:</em></strong><em> Agent memory architecture is the system that stores, retrieves, updates, and governs the information an AI agent uses across steps, conversations, and sessions, beyond what fits in the model’s context window.</em></blockquote><p>‘ <a href="https://arxiv.org/abs/2309.02427">Cognitive Architectures for Language Agents</a>,’ known as <strong>CoALA</strong>, splits agent memory into short-term working memory and long-term memory, then divides long-term memory into semantic, episodic, and procedural stores.</p><p>A 2026 paper, ‘ <a href="https://arxiv.org/pdf/2604.11364">The Missing Knowledge Layer in Cognitive Architectures for AI Agents</a>,’ points out that CoALA groups semantic and episodic memory together under “long-term” and gives neither a defined update mechanism, ownership scope, or decay behavior. Two memory types can share a category and still need completely different permission models.</p><h3>The four types of AI agent memory</h3><p><strong>Working memory</strong> holds the variables and observations an agent is actively using: the conversation so far, tool outputs from the last few steps, and anything it needs active to finish the current task.</p><p><strong>Episodic memory</strong> stores a record of past interactions and events, timestamped, so the agent has history to draw on without replaying every prior session.</p><p><strong>Semantic memory</strong> holds durable business facts, the kind that typically live in a system of record outside the model.</p><p><strong>Procedural memory</strong> captures reusable workflows and learned action patterns the agent applies across tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DGl0KH1xAztCLQ5jPjEvsQ.png" /></figure><p>Episodic and semantic memory need different retention rules for a reason: an episodic experience should decay the way a person’s memory of an ordinary day fades, while a semantic claim should get superseded when newer evidence shows up. Treat both the same way, and you either keep stale logs around forever or discard facts that are still true.</p><p>Semantic memory raises its own question. If it draws from Salesforce, NetSuite, or Snowflake, the safer pattern is reading it live, <a href="https://www.cdata.com/ai/capabilities/identity-and-access/">under the requesting user’s own permissions</a>, instead of copying it into an embedding store that has no permission model attached. An embedding doesn’t know who’s allowed to see the record it came from.</p><p>Procedural memory is lower risk by comparison. It’s the agent’s playbook, the steps it takes to close a ticket or draft a report, and it changes behavior more than data exposure.</p><h3>Short-term vs. long-term memory: Context windows, retrieval, and live data</h3><p><a href="https://www.langchain.com/blog/launching-long-term-memory-support-in-langgraph">LangGraph</a> draws a clean line between the two scopes. Short-term memory is thread-scoped and persisted by a checkpointer, tied to one conversation. Long-term memory lives in custom namespaces and can be recalled across any thread.</p><p>Stuffing more history into the context window to bridge that gap doesn’t hold up at scale. In the <a href="https://arxiv.org/pdf/2510.11967">Context-Folding approach</a>, an agent matched or beat standard <strong>ReAct</strong> baselines while running with an active context ten times smaller, by collapsing finished subtasks into summaries instead of keeping every step in view.</p><p>Short-term memory keeps the current task coherent inside the context window. Long-term memory persists across sessions, and because it needs to be retrieved rather than carried along, retrieval is exactly where access control must apply.</p><p>The practical fix is pulling live data from the source on demand, with <a href="https://www.cdata.com/blog/custom-mcp-tools-cut-token-costs">filtering pushed down to that source</a>, which keeps prompts small and answers grounded in current data.</p><h3>How governed access works across agent memory layers</h3><p>The <a href="https://genai.owasp.org/llm-top-10/">Open Worldwide Application Security Project’s (OWASP) Top 10 for LLM Applications 2026</a> moved Excessive Agency up to the third spot, and its guidance is specific: enforce identity-based, least-privilege access outside the model, not inside the prompt.</p><h3>How short-term agent memory is governed</h3><p>Session state is isolated by thread or session ID, which LangGraph handles well. That isolation protects one user’s conversation from another’s. It does nothing about a tool output written into that state that happens to include data the requesting user isn’t cleared to see. That output has to be checked against the requesting user’s permissions before it reaches the context window.</p><h3>How long-term memory retrieval enforces user-level permissions</h3><p>The <a href="https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization">Model Context Protocol (MCP) authorization specification</a> frames the server as an OAuth 2.1 resource server and the client as acting on behalf of a resource owner. In practice, that’s what passthrough authentication should mean for long-term memory: the end user’s identity, via OAuth or SAML, passes through to the source system where the source supports per-user authentication, and the source system enforces its own native permissions. It isn’t the same as forwarding a raw token and hoping the source sorts it out.</p><blockquote><strong><em>Quick definition:</em></strong><em> Privilege drift is the gradual build-up of access an agent holds beyond what any single requesting user is entitled to, usually caused by shared service accounts or long-term memory stores that keep data without the permission context it came with.</em></blockquote><h3>Audit trails and preventing privilege drift</h3><p><a href="https://www.cdata.com/blog/ai-model-audit-trail-guide">Audit logs</a> need to record the user, the agent, the memory layer, the source system, and the request behind every retrieval, and those logs need to be ready for export and review. The <a href="https://www.nist.gov/itl/ai-risk-management-framework">NIST AI RMF</a> places this under its GOVERN function, and the <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">Generative AI Profile (AI 600–1)</a> applies that to generative systems.</p><h3>Implementing governed memory in LangChain, LangGraph, CrewAI, and Google ADK</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QG6PQ5pgp3ihqqagjQyYXg.png" /></figure><p><a href="https://www.cdata.com/developers/ai/by-ai-surface/crewai/">CrewAI</a> shares <a href="https://docs.crewai.com/concepts/memory">memory across every agent</a> in a crew unless an agent is given its own. That default works fine for a single-tenant pilot. It stops working the moment two users with different permissions share a crew, which makes per-user scoping something you design in rather than something you inherit for free.</p><p><strong>Google ADK</strong> keeps the same split in different terms. Session and State cover the current conversation. MemoryService is the searchable store that can span past sessions or pull in outside data. The outside data is where the governance question lives, since MemoryService itself has no opinion about who’s allowed to see what it returns.</p><p>None of this means swapping out your orchestration layer. <a href="https://www.cdata.com/ai/">CData Connect AI</a> doesn’t replace <strong>LangGraph</strong>, <strong>CrewAI</strong>, <strong>Google ADK</strong>, or memory frameworks such as <strong>Mem0</strong>. It sits underneath them as an MCP-based connectivity and access-control layer, so a retrieval call from any of those frameworks runs with the requesting user’s actual permissions instead of a shared service account’s.</p><h3>Build a governed memory layer for your agents with CData Connect AI</h3><p>The orchestration framework owns working and episodic memory. <a href="https://www.cdata.com/ai/">CData Connect AI</a> governs how every layer reaches business data, applying inherited user permissions and OAuth or SAML passthrough no matter which layer is asking.</p><p>For a security lead weighing sign-off, that means <a href="https://www.cdata.com/blog/mcp-field-level-security">least-privilege access</a> in line with OWASP’s guidance on Excessive Agency, plus audit logs that record the user, agent, memory layer, and source system behind each retrieval, with <a href="https://www.cdata.com/ai/capabilities/governance/">metadata available for routing to your security information and event management (SIEM) system</a>.</p><p>For the team building the agent, it means retrieval stays lean. <a href="https://www.cdata.com/ai/context-engine/">Connect AI pushes filtering down</a> to the source instead of pulling full tables into the model’s context, so governance doesn’t add context or cost.</p><p><a href="https://www.cdata.com/ai/signup/">Start a free trial of CData Connect AI.</a></p><h3>Frequently asked questions</h3><p>What is agent memory architecture, and why does it matter for enterprise AI?</p><p>Agent memory architecture is the system that stores, retrieves, updates, and governs the information an AI agent uses across steps, conversations, and sessions, beyond what fits in the model’s context window. It matters because agents moving from pilot to production need that information to stay accurate and respect the requesting user’s actual permissions, not just fit in the prompt.</p><p>What’s the difference between short-term and long-term memory in an AI agent?</p><p>Short-term memory is thread-scoped and holds the active conversation, persisted by a checkpointer. Long-term memory lives in its own store, can be recalled across any thread, and must be retrieved rather than carried along, which is exactly where access control has to apply.</p><p>Where should different types of agent memory be stored?</p><p>Working memory stays in-context as session state. Episodic memory typically lives in a log store or vector database with its own retention rules. Semantic memory, when it reflects live business facts, is best read directly from the system of record, such as a CRM or data warehouse, rather than copied into an embedding store with no permission model attached.</p><p>How does an AI agent retrieve the right memory without overloading its context window?</p><p>Push filtering down to the source system instead of pulling full records into the prompt. Collapsing finished subtasks into summaries instead of keeping every step in view can shrink the active context substantially while matching baseline task performance.</p><p>How do you prevent an AI agent from accessing data a user isn’t authorized to see?</p><p>Pass the requesting user’s identity through to the source system via OAuth or SAML and let that system enforce its own native permissions, rather than routing requests through a shared service account. Log the user, agent, memory layer, source system, and request behind every retrieval so access can be audited, not just assumed.</p><p>What are the most common mistakes in agent memory architecture design?</p><p>The two that show up most often: treating episodic and semantic memory as interchangeable when they need different decay and update rules, and copying business data into an embedding store that carries no permission context from the source system it came from.</p><p>How should memory be scoped and isolated in a multi-agent system like LangGraph?</p><p>LangGraph isolates short-term state by thread or session ID and keeps long-term memory in custom namespaces. Namespaces separate data between agents, but they don’t check source-system permissions on their own, so scoping by user identity still has to be built in rather than assumed.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/agent-memory-architecture-governed-access"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a7b14833a0d6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/agent-memory-architecture-short-long-term-governed-access-a7b14833a0d6">Agent Memory Architecture: Short- &amp; Long-Term Governed Access</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP Gateway Security Tools Comparison for Enterprise (2026)]]></title>
            <link>https://medium.com/cdata-software/mcp-gateway-security-tools-comparison-for-enterprise-2026-c44eb43853d4?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/c44eb43853d4</guid>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 17:11:51 GMT</pubDate>
            <atom:updated>2026-11-05T17:11:51.151Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*3POnG9hLjmKn91mO" /></figure><p>Security for AI agents works best when it’s built in from the first deployment. Once agents hold credentials to production systems, every missing control becomes a path to live data. The <a href="https://modelcontextprotocol.io/">Model Context Protocol</a> (MCP) standardizes how agents reach those systems, yet in the current <a href="https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization">MCP authorization spec</a>, authorization is optional, and identity, audit, and tool vetting fall to each team that deploys it.</p><p>An MCP gateway closes that gap by centrally brokering, authenticating, monitoring, and protecting the calls AI agents make to enterprise data and tools. The NSA’s <a href="https://www.nsa.gov/Portals/75/documents/Cybersecurity/CSI_MCP_SECURITY.pdf">first MCP security guidance</a>, published in May 2026, found that MCP adoption has outpaced its security model.</p><h3>Why enterprises need an MCP gateway</h3><p>Most MCP rollouts start with one server per source, each with its own credentials, tools, and logs. Agents then chain calls across those servers, so untrusted text from one source can trigger an action in another.</p><p>Invariant Labs showed how that plays out with the <a href="https://invariantlabs.ai/blog/mcp-github-vulnerability">GitHub MCP server</a>. A malicious public issue coerced an agent into reading private repositories and leaking the data in a public pull request. Nothing narrowed the agent’s token, which reached every repository.</p><p>Zero Trust networks and API gateways verify who called an endpoint, yet they can’t see which MCP tool an agent picked or whether the person behind it should see the rows returned. An MCP gateway adds that agent-aware layer to the controls you already run.</p><blockquote>Quick definition: defense-in-depth</blockquote><blockquote><em>Defense-in-depth stacks independent controls, such as identity, authorization, and audit, so one failure doesn’t expose the whole system.</em></blockquote><h3>Core security features of MCP gateways</h3><p>A modern MCP gateway moves security from scattered, per-server settings to one managed platform. Expect these baseline MCP security controls.</p><h3>Authentication and authorization controls</h3><p>For remote servers, the MCP spec builds on OAuth 2.1 with Proof Key for Code Exchange (PKCE), and OpenID Connect (OIDC) or mutual TLS (mTLS) can strengthen client and server identity. Servers must accept only <a href="https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/security-considerations">tokens issued for them</a> and must never pass a client’s token through to upstream APIs.</p><blockquote>Quick definition: capability-scoped tokens</blockquote><blockquote><em>Capability-scoped tokens are short-lived credentials that bind an identity to a specific set of roles and permissions.</em></blockquote><h3>Identity-first security and role-based access</h3><p>Identity-first security anchors every agent and tool action to an authenticated user or service, enforced at the gateway. Role-based access control (RBAC) and attribute-based access control (ABAC) bind privileges to scoped tokens, and the audit trail records each action by user, agent, and session.</p><h3>Protocol validation and runtime defenses</h3><p>A gateway should validate every message against the MCP spec and reject anything malformed. Runtime defenses add injection filters, rate limiting, bot detection, and schema checks on tool parameters.</p><h3>Key threats addressed by MCP gateways</h3><p>The <a href="https://owasp.org/www-project-mcp-top-10/">OWASP MCP Top 10</a> gives security teams a shared vocabulary for MCP risks, and any gateway you shortlist should defend against at least three attack types from it.</p><h3>Prompt injection and command injection attacks</h3><blockquote>Quick definition: prompt injection</blockquote><blockquote><em>Prompt injection is when a threat actor manipulates an AI agent’s instructions, tricking the model into running malicious or unauthorized commands.</em></blockquote><p>In a Cato CTRL <a href="https://www.catonetworks.com/blog/cato-ctrl-poc-attack-targeting-atlassians-mcp/">proof of concept</a>, a malicious Jira Service Management ticket ran with internal privileges when an engineer used an AI action through Atlassian’s MCP server. In April 2026, OX Security reported a <a href="https://thehackernews.com/2026/04/anthropic-mcp-design-vulnerability.html">command execution flaw</a> in official MCP SDKs that passed STDIO configuration to the host shell unsanitized. Input filtering limits prompt injection, while schema validation and keeping user input away from shell calls limit command injection.</p><h3>Tool poisoning and supply chain risks</h3><blockquote>Quick definition: tool poisoning</blockquote><blockquote><em>Tool poisoning is the compromise of an MCP tool or package to introduce malicious actions or escalate privileges, often through hidden instructions in tool descriptions.</em></blockquote><p>A server that changes its tool descriptions after approval can redirect an agent without tripping an alert. A sound supply chain process has four steps.</p><ol><li>Catalog every MCP server, tool, and package in use.</li><li>Sign artifacts and pin versions so nothing changes silently.</li><li>Re-review tool descriptions on every update.</li><li>Retire servers that lose an active maintainer.</li></ol><h3>Credential misuse and data exfiltration</h3><blockquote>Quick definition: data exfiltration</blockquote><blockquote><em>Data exfiltration is the unauthorized transfer of data out of the enterprise through compromised servers, manipulated agents, or leaked responses.</em></blockquote><p>In the GitHub case, one overprivileged token carried private repository data into a public pull request, and four controls close that path.</p><ul><li>Issue short-lived, audience-bound tokens.</li><li>Run queries on each user’s own source credentials where possible.</li><li>Mask personally identifiable information (PII) and secrets in responses before they reach the model.</li><li>Alert on unusual data volumes or new tool sequences.</li></ul><h3>Architectural best practices for MCP gateway deployment</h3><p>Run one gateway as the control plane for all AI traffic. A centralized control plane is the single layer that enforces policy and collects telemetry across every agent and MCP server. It allowlists servers, orchestrates sessions, filters which tools each agent sees, and logs every call, which brings MCP sprawl under one set of controls.</p><p>Use OAuth 2.1 and OIDC for user and agent sessions, mTLS between gateway and servers, and a managed vault with scheduled key rotation.</p><p>Cloud-native gateways deploy faster, while regulated enterprises often need on-premises or hybrid options. Test latency, throughput, high availability, and disaster recovery for each, and keep toolsets lean.</p><h3>Essential security controls for enterprise MCP gateways</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZVwJ4Bnxr0tLoV1RPZDQCg.png" /></figure><p>Review new tools before approval, re-check permissions on a schedule, and route PII alerts into your observability stack.</p><h3>Observability and incident response</h3><p>An audit trail records who did what, when, and through which tool. A security information and event management (SIEM) system correlates those logs into alerts, and forensic tracebacks rebuild incidents from them.</p><h3>Centralized logging and immutable audit trails</h3><p>An immutable audit trail is a log record that can’t be changed, deleted, or overwritten. Aggregate logs from every gateway component into your SIEM, with retention and access controls set by regulation.</p><h3>SIEM correlation and telemetry integration</h3><p>Monitor failed authentication, denied tool calls, anomalous tool usage, and data loss prevention alerts. OpenTelemetry tracing follows one request across agents, tools, and sessions.</p><h3>Incident playbooks and forensic investigations</h3><ol><li>Contain the affected user, agent, or server.</li><li>Revoke tokens and rotate credentials.</li><li>Isolate the suspect tool or server.</li><li>Preserve logs and session fingerprints.</li><li>Escalate and reconstruct the timeline.</li></ol><p>Rehearse the playbook in regular tabletop exercises so each team knows its role before an incident.</p><h3>Automating SecOps for MCP security</h3><p>Static validation runs pre-deployment checks for vulnerabilities and misconfigurations in MCP tools, so build it into CI/CD and log every approval.</p><p>Runtime behavioral detection catches zero-day techniques that static checks miss. Flag unexpected tool calls, unauthorized data flows, and shifts in prompt patterns, and alert the agent’s owner.</p><p>Configuration drift remediation corrects systems that deviate from a secure baseline by detecting drift, alerting the owner, remediating, and reporting. Infrastructure as code (IaC) keeps that baseline versioned.</p><h3>Cataloging and managing MCP clients, servers, and tools</h3><p>Keep a private registry of every approved MCP client, server, and tool, and run automated discovery to find shadow deployments. The NSA guidance recommends <a href="https://www.cdata.com/blog/mcp-security-enterprise-nsa-advisory">regular scans</a> for unauthorized or unauthenticated MCP servers.</p><p>Onboard through role-based approvals with a named owner per server, and decommission by revoking credentials and archiving logs.</p><h3>Governance and compliance considerations</h3><p>For AI governance, an MCP gateway gives auditors one place to verify identity tracking, access restrictions, and logging.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ar97edFBxt2Kjw0dQ_ek0Q.png" /></figure><p>RBAC assigns permissions by role, ABAC uses attributes of the user, resource, and context, and relationship-based access control (ReBAC) uses relationships among users and resources. Many teams pair RBAC with ABAC, written as policy-as-code so every change is reviewed and tested.</p><h3>Comparing leading MCP gateway security solutions</h3><p>MCP gateway security tools fall into a few architectural groups, and the right choice depends on what the gateway can see.</p><h3>CData Connect AI Gateway</h3><p><a href="https://www.cdata.com/ai/mcp-gateway/">CData Connect AI Gateway</a> is an MCP gateway built on a governed data layer. It connects agents to hundreds of enterprise systems and to third-party MCP servers through the same gateway, so every tool sits behind one set of controls.</p><p>Because it enforces security at the data layer as well as the traffic layer, it knows whose permissions apply at the source, along with which tool was called. Connect AI Gateway lets you:</p><ul><li>Sign users in through OAuth or SAML and run queries on their own source credentials where sources support it.</li><li>Scope tools by workspace and toolkit, with calls denied and logged when a check can’t be evaluated.</li><li>Keep credentials vaulted and injected server-side, never exposed to the agent.</li><li>Stream one audit trail of every tool call to your SIEM.</li></ul><p>CData holds SOC 2 Type II and ISO 27001 certifications and is GDPR compliant (see the <a href="https://www.cdata.com/ai/security/">Connect AI security page</a>).</p><h3>Key selection criteria for enterprise leaders</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UmKFI-SeWoXxueaUgrwESw.png" /></figure><p>Score enterprise MCP gateway solutions on coverage, compliance, ease of management, performance, deployment options, extensibility, identity integration, and SecOps automation. Ask every vendor these questions.</p><ul><li>Does the gateway enforce permissions at the source, or only at the tool?</li><li>Can third-party MCP servers sit behind the same identity, policy, and audit?</li><li>Do denied or unevaluated calls fail closed?</li><li>Can agents ever see upstream credentials?</li></ul><h3>Next steps for enterprise leaders implementing MCP gateway security</h3><p>Each phase ends with a success measure.</p><ol><li>Discover every MCP client, server, and tool until no unknown servers remain.</li><li>Deploy one gateway and retire direct agent-to-server connections.</li><li>Enforce SSO and per-user credentials so every call maps to a person or service.</li><li>Turn on runtime defenses and confirm blocked attempts appear in logs.</li><li>Stream audit events to your SIEM and test that alerts fire.</li><li>Govern with policy-as-code and scheduled reviews so audit evidence is on hand.</li></ol><p>For a deeper checklist, download the <a href="https://www.cdata.com/media/jioaw0lw/cdatasecuritybestpracticesguide2026.pdf">security best practices guide</a>.</p><h3>Secure every AI agent call with CData Connect AI Gateway</h3><p><a href="https://www.cdata.com/ai/">CData Connect AI Gateway</a> puts identity, tool-level policy, and audit in front of every MCP server your agents use, CData-built or third-party. Every call carries a named identity and lands in one audit trail your security team reviews, so agents reach live enterprise data without opening new gaps.</p><p><a href="https://www.cdata.com/ai/signup/">Start a free trial</a> and route your first agent through one governed gateway.</p><h3>Frequently asked questions</h3><p>What is an MCP gateway and how does it differ from traditional API gateways?</p><p>An MCP gateway is built for AI agents and enforces identity and policy controls around the Model Context Protocol. API gateways handle generic API traffic and lack agent-specific controls such as tool-level authorization.</p><p>How does an MCP gateway improve security for AI agent deployments?</p><p>It centralizes authentication, access control, protocol validation, and threat detection, which reduces data leakage, supply chain, and prompt injection risk.</p><p>What are the most common MCP-specific threats enterprises should prepare for?</p><p>Prompt and command injection, tool poisoning, supply chain compromise, credential misuse, and data exfiltration lead the list.</p><p>How should identity and access be managed within an MCP gateway?</p><p>Use OAuth 2.1, OIDC, or mTLS with granular role-based access controls for users, agents, and tools, and run queries on each user’s own source credentials where possible.</p><p>What observability and incident response capabilities should a secure MCP gateway provide?</p><p>It should offer centralized logging, immutable audit trails, SIEM integration, and fast incident response.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/mcp-gateway-security-enterprise-leaders"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c44eb43853d4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/mcp-gateway-security-tools-comparison-for-enterprise-2026-c44eb43853d4">MCP Gateway Security Tools Comparison for Enterprise (2026)</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Connectors in ETL: 2026 Guide to Maintaining Connectors]]></title>
            <link>https://medium.com/cdata-software/data-connectors-in-etl-2026-guide-to-maintaining-connectors-a674a60d591a?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/a674a60d591a</guid>
            <category><![CDATA[data-integration]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[data-architecture]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[etl]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 16:56:39 GMT</pubDate>
            <atom:updated>2026-11-05T16:56:39.572Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*rKhpZoVUJU37KvBB" /></figure><p>Connector maintenance is a structural tax on every data team. Source application programming interfaces (APIs) change, schemas shift, and tokens expire, and each event lands on the engineers who own the pipelines. Managed connectors move that upkeep off the team’s payroll, and the trade-off against building in-house comes down to source volatility and engineering capacity. With <a href="https://www.cdata.com/sync/">CData Sync</a>, hundreds of pre-built, actively maintained connectors absorb that work.</p><h3>What are data connectors in ETL?</h3><p>A <strong>data connector</strong> is a reusable software component that establishes and manages the connection between a <strong>source system</strong> and a <strong>destination</strong>, so data moves automatically through an extract, transform, load (ETL) or extract, load, transform (ELT) pipeline. In ETL, data connectors handle the extract and load phases, retrieving records from a source system and pushing them into a destination such as a data warehouse, while transformation happens separately in the tool or the destination, as <a href="https://www.ibm.com/think/topics/etl">IBM</a> describes.</p><p>ETL transforms data before loading, and ELT loads raw data first and transforms it in the destination. Connectors perform the same extract and load work in both patterns, so connector maintenance practices apply equally to ETL and ELT. CData’s overview of <a href="https://www.cdata.com/blog/what-are-data-connectors">data connectors</a> covers the main connector types and how each fit into a pipeline.</p><h3>Common types of ETL data connectors</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*scevOVschFaLAw1C4VflTQ.png" /></figure><h3>Failure modes and root causes behind broken ETL connectors</h3><p>Two root causes drive most connector breakage: source API version changes and schema drift. <a href="https://learn.microsoft.com/en-us/azure/data-factory/concepts-data-flow-schema-drift">Microsoft Learn</a> explains why rigid source mappings fail when schemas change.</p><p><strong>Schema drift</strong> is when the structure of source data, including new, renamed, dropped, or retyped fields, changes unexpectedly in ways the pipeline’s design did not anticipate. The most dangerous form fails silently. A renamed source field, for example, can let a sync finish green while a revenue field loads null for days, leaving dashboards wrong until someone notices.</p><p>API changes cause the second failure mode. SaaS vendors add, rename, or reformat fields in new API versions, and connectors with mappings tied to specific source names break when upstream sources evolve without notice. Vendors such as Salesforce publish new API versions on a regular cadence, so mappings need regular review. Ad hoc drift handling stops scaling once several sources feed the same warehouse, because each added source multiplies the odds of an undetected drift event.</p><h3>How to monitor connector health and detect issues early</h3><p>Monitoring tells a team the pipeline ran, while observability tells whether it produced correct, complete, and fresh data. A pipeline can report success and still deliver zero rows, which is why the job-ran signal alone cannot establish trust in a connector.</p><p>Data freshness is often the first signal that a connector silently stopped. CData Sync detects source schema changes on each run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qgVNC2qULgJKSu61dTtnsA.png" /></figure><h3>A step-by-step workflow to update and test connectors</h3><p>When an API change or drift event forces a fix, a repeatable workflow keeps the update predictable: detect the trigger, review the changelog, update mappings, test in staging, and verify after deploying.</p><ul><li><strong>Detect and triage:</strong> Confirm the trigger, whether a deprecation notice, drift alert, or failed run, and rank affected fields by criticality. Critical fields get hard alerts, and optional fields get soft warnings.</li><li><strong>Review the changelog:</strong> Read the source’s deprecation notice to identify which fields, endpoints, or parameters changed. Breaking changes typically remove or rename one of the three.</li><li><strong>Update mappings and auth:</strong> Revise field mappings and handle authentication changes explicitly, since expired tokens and <strong>Open Authorization (OAuth)</strong> grant updates commonly break connectors.</li><li><strong>Test in staging:</strong> Run contract tests and <strong>canary runs</strong>. A <strong>contract test</strong> validates that incoming data matches an agreed schema before a change deploys, which catches drift in continuous integration (CI) or staging instead of production.</li><li><strong>Deploy and verify:</strong> Confirm freshness and row counts return to baseline and treat pre-deploy testing as mandatory, since a version change can alter every mapping at once.</li></ul><p>Vendor <strong>Sunset headers</strong> and deprecation dates let teams schedule updates ahead of removal instead of reacting after breakage. Managed vendors publish the same information in release notes, as in the <a href="https://www.cdata.com/blog/sync-q3-2025-release-open-tables-cdc-reverse-etl">CData Sync Q3 2025 release post</a>.</p><h3>Should ETL teams build or buy connectors?</h3><p>The build is the cheap part, and the bulk of the work is maintenance. Six months to build one performant connector, and up to two years to cover complex sources such as Workday or NetSuite, according to <a href="https://www.cdata.com/blog/build-vs-buy-saas-integrations">CData’s build vs. buy analysis</a>. With CData Sync, connection-based pricing keeps costs predictable as data volume grows.</p><p>On the build side, every source API or schema change forces the internal team to re-engineer the integration, while managed vendors provide connectors that stay current.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bcAul4wfGkJ5zmNLgPGAbQ.png" /></figure><p>CData Sync represents the buy path, with hundreds of pre-built, actively maintained connectors. One example is the <a href="https://www.cdata.com/kb/articles/workday-sync.rst">Workday integration</a>. A hybrid model buys managed connectors for volatile third-party SaaS sources and builds for a small set of proprietary internal systems. CData’s guides to <a href="https://www.cdata.com/blog/choose-etl-with-dbt-2026">choosing ETL with dbt</a> and <a href="https://www.cdata.com/blog/best-real-time-etl-tools-cloud-on-prem-sync">real-time ETL tools</a> cover adjacent decisions.</p><h3>Maintain reliable ETL connectors with CData Sync</h3><p>With <a href="https://www.cdata.com/sync/">CData Sync</a>, schema drift handling, authentication updates, and version compatibility are the vendor’s job, so engineers stop patching deprecated endpoints connector by connector. <a href="https://www.cdata.com/sync/trial/">Start a free trial of CData Sync</a> to see managed connectors at work.</p><h3>Frequently asked questions</h3><p>What are the different types of data connectors in ETL?</p><p>Common types include database, API or web service, file-based, cloud SaaS application, and reverse ETL connectors. Each type links a specific kind of source or destination to the pipeline.</p><p>How do you update an ETL connector when a source API changes?</p><p>Review the changelog, update field mappings and authentication, test in staging with contract and canary runs, then deploy and verify that freshness and row counts return to baseline.</p><p>What are the most common reasons ETL data connectors fail?</p><p>Source API version changes, schema drift, and expired or changed authentication cause most failures. Rigid field mappings turn each of these into a broken or silently incorrect load.</p><p>What tools do you need before you can maintain data connectors in ETL?</p><p>Maintaining data connectors in ETL requires freshness and row-volume monitoring, schema drift detection against a stored baseline, a staging environment, and contract tests.</p><p>Should you build your own ETL connectors or use managed connectors?</p><p>Build for niche or internal systems with stable interfaces and buy for frequently changing SaaS sources. Many teams combine both.</p><p>How often should ETL connectors be reviewed and updated?</p><p>Review connectors after every source API version release, and at minimum quarterly.</p><p>What is the difference between ETL and ELT, and does it affect how connectors are maintained?</p><p>ETL transforms data before loading, and ELT loads first and transforms in the destination. Connectors handle extract and load in both patterns, so maintenance practices apply equally.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/data-connectors-etl-maintaining-connectors-2026"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a674a60d591a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/data-connectors-in-etl-2026-guide-to-maintaining-connectors-a674a60d591a">Data Connectors in ETL: 2026 Guide to Maintaining Connectors</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[11 Essential AI Data Governance Tools to Secure & Scale in 2026]]></title>
            <link>https://medium.com/cdata-software/11-essential-ai-data-governance-tools-to-secure-scale-in-2026-4f188d31e16d?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/4f188d31e16d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 16:44:22 GMT</pubDate>
            <atom:updated>2026-11-05T16:44:22.368Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*Nee7YBRzwQoBsooY" /></figure><p>An AI assistant connects to a customer database using a shared service account. For InfoSec teams, this makes it difficult to identify who requested the data or verify whether they had permission to access it. Using the company’s existing identity system allows access to be tied to individual users, with permissions enforced at the data source.</p><h3>What is AI data governance</h3><p>AI data governance is the set of policies and controls that decide which enterprise data an AI system can reach, what it can do with it, and how each request is tied to a person. Traditional data governance often relies on periodic access reviews, while AI agents require permissions to be checked each time they access or act on live data.</p><h3>Why AI data governance matters in 2026</h3><p>When five teams each connect an agent with its own credential, security inherits five unreviewed access paths and no central enforcement point. Each connection introduces potential data exposure risks.</p><p>Auditors expect a trail of what each AI interaction sent, returned, and decided, and who authorized it. The <strong>EU AI Act</strong> introduces risk-based requirements, with obligations for <strong>Annex III high risk AI systems</strong> taking effect from December 2, 2027. State privacy and industry regulations also govern how AI processes personal data across cloud environments. CData’s guide on <a href="https://www.cdata.com/blog/act-now-on-ai-governance">why organizations should act now on AI governance</a> explains the importance of preparing for these requirements.</p><h3>What makes a tool “enterprise ready” for AI governance</h3><p>An enterprise AI governance tool should offer verified security controls, user level permissions, audit logging, and workspace isolation. Access should follow the <strong>national institute of standards and technology’s (NIST)</strong> principle of least-privilege, giving users and agents only the permissions needed for their tasks. <a href="https://www.cdata.com/ai/">CData Connect AI</a> helps enforce these controls when AI agents access enterprise data sources.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GBwMoUowhtjVloFD2O13nA.png" /></figure><h3>SOC 2 Type II and ISO 27001 certifications</h3><p><strong>SOC 2</strong> evaluates controls against the <strong>Trust Services Criteria</strong> established by the <strong>American Institute of Certified Public Accountants (AICPA)</strong>, covering security, availability, processing integrity, confidentiality, and privacy. A SOC 2 Type II report evaluates how effectively controls operated over roughly six to twelve months, providing stronger evidence of ongoing compliance than a Type I report, which assesses controls at a specific point in time.</p><p>Security is the only mandatory category, so check which additional criteria are covered. <strong>ISO/IEC 27001</strong> certifies an organization’s information security management system. For both, verify that the scope includes the product being evaluated. CData outlines its certifications and audit coverage in the <a href="https://www.cdata.com/ai/security/">Connect AI security and compliance page</a>.</p><h3>OAuth/SAML identity passthrough and least privilege access</h3><p><strong>Identity passthrough</strong> allows AI agents to access data using the authenticated user’s identity through <strong>open authorization (OAuth)</strong> or <strong>security assertion markup language (SAML)</strong>, so source systems can enforce individual permissions. Shared credentials weaken this control by masking individual users. The <a href="https://www.cdata.com/blog/mcp-server-best-practices-2026">model context protocol (MCP)</a> provides a framework for authorization, with each implementation responsible for enforcing access permissions. Connect AI supports <a href="https://www.cdata.com/ai/capabilities/identity-and-access/">per user identity passthrough</a> to preserve user permissions across connections. Its guide to <a href="https://www.cdata.com/blog/managing-data-permissions-enterprise-ai-agents">managing data permissions for enterprise AI agents</a> explains how these controls apply to agent access.</p><h3>Audit logging and SIEM exportability</h3><p>Every AI request should have an audit record that identifies the user and can be exported to the organization’s <strong>security information and event management (SIEM)</strong> system. Audit logging and monitoring support SOC 2 requirements and help organizations demonstrate compliance with regulations such as the <strong>health insurance portability and accountability act (HIPAA)</strong> and the <strong>Sarbanes Oxley act (SOX)</strong>.</p><h3>Workspace isolation and pilot to production scalability</h3><p>Workspace isolation partitions AI data access into separately governed workspaces, so credentials, permissions, and audit scopes stay within one team or environment. Governance needs to be centralized as AI adoption expands beyond individual pilots to multiple teams. Workspaces in Connect AI can scope each team’s tools once and carry governance across AI platforms.</p><h3>The 11 essential AI data governance tools for 2026</h3><p>AI data governance involves cataloging, data lineage, identity management, data loss prevention (DLP), and connectivity. These capabilities help maintain data quality, reduce silos, and enforce access policies. The certification information below reflects what vendors published on their trust pages as of October 2026.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DJTs07krF_5gmz5smETaww.png" /></figure><h3>1. Connect AI</h3><p>Connect AI provides governed access to enterprise data for AI assistants and agents. It supports connectivity through the MCP, ODBC, JDBC, REST, and OData. Its <a href="https://www.cdata.com/ai/capabilities/governance/">governance capabilities</a> include user level permissions, access policies, and audit logging to help enforce least privilege.</p><p><strong>Connect AI Gateway</strong>, currently in early access, provides a centralized control point for model routing, tool access, and enterprise data connections. It applies access policies throughout agent interactions, enforces permissions at the data source, and records activity for auditing.</p><h3>2. Alation</h3><p><strong>Alation</strong> is a data catalog that helps organizations discover and manage data assets through a searchable inventory of definitions, ownership details, and governance policies. Data stewards can use it to identify trusted datasets and determine which assets are suitable for AI use.</p><h3>3. SailPoint Identity Security Cloud</h3><p><strong>SailPoint</strong> helps organizations manage identity permissions and enforce least privilege across AI connected systems. CISOs can use it to govern access for human and nonhuman identities, review entitlements, and track access approvals. These capabilities align with the NIST access control requirements, including AC-2 for account management and AC-3 for access enforcement.</p><h3>4. Monte Carlo</h3><p>Data lineage maps where data originates and how it changes on the way to a report, model, or agent, which a catalog or data model alone can’t show. <strong>Monte Carlo</strong> builds field level lineage for regulatory traceability and AI training data, so compliance officers can trace an AI output to its source during an audit.</p><h3>5. BigID</h3><p>BigID finds and classifies personally identifiable information (PII), financial, and health data and marks what should never reach a prompt or a training set. Privacy leaders use that inventory as SOC 2 Confidentiality evidence.</p><h3>6. Kong AI Gateway</h3><p><strong>Kong AI Gateway</strong> provides centralized control over AI model requests, allowing platform teams to enforce policies, monitor usage, and manage costs. It also supports prompt security controls and PII protection.</p><h3>7. Microsoft Sentinel</h3><p><strong>Microsoft Sentinel</strong> is an SIEM platform that helps security teams monitor AI activity alongside other security events. It can collect logs from AI applications and connected systems to detect suspicious behavior and support investigations. These capabilities align with the NIST audit and accountability controls.</p><h3>8. Credo AI</h3><p><strong>Credo AI</strong> helps organizations automate AI risk assessments and compliance tracking using policy packs aligned with the voluntary <strong>NIST AI risk management framework</strong>. The framework organizes AI risk management into four functions, <strong>Govern</strong>, <strong>Map</strong>, <strong>Measure</strong>, and <strong>Manage</strong>. Organizations can use it to assess and document AI risks, although NIST does not certify compliance with the framework.</p><h3>9. Ataccama ONE</h3><p><strong>Ataccama ONE</strong> helps data engineering teams improve the quality of data used by AI systems through automated profiling, validation, and continuous monitoring. These capabilities help identify inaccurate or inconsistent data that could lead to unreliable AI outputs.</p><h3>10. Databricks Unity Catalog</h3><p><strong>Databricks Unity Catalog</strong> provides centralized governance for data assets and AI models within Databricks, including tables, files, machine learning models, and features. It supports access controls, audit logging, data lineage, and quality monitoring, helping teams apply consistent policies across their data and AI workloads. Assets outside the Databricks environment may require additional governance tools.</p><h3>11. Immuta</h3><p>Data masking swaps sensitive values for realistic substitutes, either in a sanitized copy (static) or at request time based on the user, leaving the source unchanged (dynamic). <strong>Immuta</strong> applies dynamic masking through policy, so privacy engineers can give AI realistic data without exposing PII, which supports GDPR compliance.</p><h3>Governed vs. ungoverned AI data access: the cost of inaction</h3><p>Governed access verifies user identities and enforces permissions for every AI request in real time through a centralized framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fD4Q_QRDCh2Vj2ZBCc6lJA.png" /></figure><p>With Connect AI, you can replace an ungoverned direct connection with a governed, auditable one. Check here on how to <a href="https://www.cdata.com/ai/solutions/ai-governance/">govern every AI tool from one control plane</a>.</p><h3>How to choose the right AI data governance tools for your stack</h3><p>Evaluate AI governance tools for least privilege, integration with existing identity and access management (IAM), SOC 2 Type II coverage, identity passthrough, exportable audit logs, and workspace isolation. Use the NIST AI RMF to identify gaps across connectivity, cataloging, lineage, DLP, and monitoring. For a side-by-side comparison, see our guide to <a href="https://www.cdata.com/blog/top-tools-governed-llm-data-access">tools for governed LLM data access</a>.</p><h3>Say “yes” to governed AI data access with CData Connect AI</h3><p>With <a href="https://www.cdata.com/ai/">CData Connect AI</a>, your next AI access request reaches security review with identity passthrough and SIEM exportable audit logs in place. It supplies the access governance MCP leaves to implementers, with the audit trails regulators expect. Evaluate CData Connect AI against your IAM before your next AI pilot review or <a href="https://www.cdata.com/ai/signup/">start a free trial</a>.</p><h3>Frequently asked questions</h3><p>What is AI data governance, and why does it matter?</p><p>AI data governance defines what data AI can access, what actions it can perform, and how activity is tracked. It helps organizations maintain secure, auditable AI access.</p><p>What features matter most in AI data governance tools?</p><p>Look for identity passthrough, per user permissions, centralized audit logs, workspace isolation, and relevant certifications such as SOC 2 Type II.</p><p>How is AI data governance different from traditional data governance?</p><p>AI governance requires access checks and audit logging for every agent request, including actions performed on live data.</p><p>What are the biggest risks of connecting AI tools to enterprise data?</p><p>Key risks include overly broad permissions, sensitive data exposure, shared credentials, and incomplete audit logs.</p><p>Can AI data governance tools work across Copilot, ChatGPT, and Claude without vendor lock-in?</p><p>Yes. Platforms like CData Connect AI use MCP to provide governed data access across supported AI clients, including Claude, Copilot, and ChatGPT.</p><p>What are the five pillars of data governance for AI?</p><p>Data quality, security and access, privacy and compliance, ownership and stewardship, and data lineage.</p><p>How do AI data governance tools compare on SOC 2 and ISO 27001?</p><p>Many vendors offer SOC 2 Type II and ISO/IEC 27001 compliance documentation. Verify that their audit reports cover the specific product being evaluated.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/ai-data-governance-tools-2026"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f188d31e16d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/11-essential-ai-data-governance-tools-to-secure-scale-in-2026-4f188d31e16d">11 Essential AI Data Governance Tools to Secure &amp; Scale in 2026</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP Gateway vs MCP Server: Performance, Scale, and Control]]></title>
            <link>https://medium.com/cdata-software/mcp-gateway-vs-mcp-server-performance-scale-and-control-d6a54e86e4cb?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/d6a54e86e4cb</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 15:24:10 GMT</pubDate>
            <atom:updated>2026-11-05T15:24:10.425Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*3FhGzs05idxbNutc" /></figure><p><a href="https://www.cdata.com/blog/what-is-mcp">Model Context Protocol (MCP)</a> helps enterprises connect AI agents to business data, applications, and tools. A single MCP server can expose multiple capabilities, but organizations may also deploy separate servers for specialized workflows, custom tools, or team-specific requirements. As these deployments grow, managing authentication, access policies, monitoring, and performance across multiple servers becomes more complex. An MCP gateway is a centralized layer that routes requests, enforces security policies, and monitors activity across those servers.</p><h3>Understanding MCP server and MCP gateway roles</h3><p>An <strong>MCP server</strong> is a backend service that exposes tools, resources, and prompts to AI applications. It lets an agent retrieve information or run actions in a connected system.</p><p>An <strong>MCP gateway</strong> is a reverse proxy for AI agents: it forwards each request to the right server. It centralizes authentication, authorization, routing, request limits (rate limiting), and monitoring, depending on the product.</p><p>For example, an assistant needs sales, finance, and support data. Without a gateway, it signs in to three servers separately. With a gateway, it uses one endpoint. <strong>MCP servers provide functionality and context. MCP gateways provide centralized authentication, governance, and control</strong>.</p><h3>Key performance differences</h3><p>MCP servers process tool requests directly, while MCP gateways add a layer for routing, traffic management, and policy enforcement. The gateway’s added cost grows with each control you enable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GYDjWOkOgx87VXwJhPzS6A.png" /></figure><p>Under the <a href="https://blog.modelcontextprotocol.io/posts/2026-07-28/">current MCP specification</a>, each request names its method and tool in the header, so a gateway can route it without reading the full message. Throughput, latency, and cost differences vary by deployment, so measure them on your own workloads.</p><h3>Scalability and deployment considerations</h3><p>A server scales by running more copies, called instances, behind a load balancer, which spreads requests across them. Gateways address a different kind of growth: more servers, more teams, and more environments.</p><p><strong>MCP server</strong></p><ul><li>Handles any request on any instance, so many assistant conversations run at once and instances share no session data</li><li>Needs deployment, patching, and monitoring for each instance</li></ul><p><strong>MCP gateway</strong></p><p>Gateway capabilities vary by product:</p><ul><li>Gives one list of servers across cloud and on-premises environments. A <a href="https://docs.cloud.cdata.com/en/MCP-Gateway">tool server</a> groups several MCP servers, including third-party ones, behind one address, the agent URL</li><li>Fails over and balances load across server instances. A <a href="https://docs.cloud.cdata.com/en/Private-Cloud-Connector">Private Cloud Connector</a> bridges on-premises data to the gateway and can run as multiple instances, so requests reroute if one fails</li><li>Standardizes deployment, since a managed gateway removes hosting work and a self-hosted one needs its own deployment and upkeep</li></ul><p>A gateway doesn’t remove the work of deploying the servers behind it.</p><h3>Security and access control features</h3><p>MCP servers enforce security controls within their individual deployments, while MCP gateways provide a centralized layer for managing access across multiple servers. Regulated organizations need that consistency in policies, auditability, and compliance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*msxtXNh1x2IF3-Z_G2coGw.png" /></figure><p>A gateway acts as a security guardrail in front of the servers. It complements the permissions that backend systems enforce, and it’s itself a component to secure and monitor.</p><h3>Management and operational oversight</h3><p>Without a gateway, operators must aggregate logs and metrics across individual MCP servers to maintain consistent oversight. Organizations should restrict direct server access, retain backend logs, and monitor gateway availability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XmtrxcHYDdyAqPomFVCO6w.png" /></figure><h3>Pros and cons of MCP gateway and MCP server</h3><p>A single point of failure (SPOF) is a component whose failure stops the whole system when no redundant copy exists. Centralized governance means access policy, auditing, and controls are set in one place for every server and team.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OpFg01_Nk2KD5lDBHNi4gg.png" /></figure><p>Direct connections fit a few servers owned by one team. A gateway fits many servers, several teams, or audit requirements, as long as you run it with failover and route all access through it.</p><h3>Choosing between MCP gateway and MCP server</h3><p>The choice depends on who manages access and how many teams share the servers.</p><ul><li>Choose an MCP server when individual teams can manage their own authentication, security, and monitoring</li><li>Choose an MCP gateway for cross-team coordination, centralized OAuth or SSO, consistent access policies, and unified audit trails</li></ul><p><strong>Check these three questions:</strong></p><ol><li><strong>Shared access</strong>: Are multiple teams integrating AI with overlapping data sources or tools?</li><li><strong>Governance</strong>: Do compliance requirements demand centralized authentication, authorization, and auditing?</li><li><strong>Growth</strong>: Do server and team counts keep growing, or must agents discover tools across servers?</li></ol><p>If any answer is yes, use a gateway-first or hybrid design.</p><h3>The managed option: CData Connect AI Gateway</h3><p><a href="https://www.cdata.com/ai/">CData Connect AI Gateway</a> combines an MCP server and an MCP gateway in one managed platform. As a server, it gives agents governed access to hundreds of enterprise data sources through one endpoint. As a gateway, it also connects third-party MCP servers, so administrators can offer their approved tools alongside its own.</p><p>With Connect AI Gateway, you can:</p><ul><li>Sign users in through your SSO provider and use <a href="https://docs.cloud.cdata.com/en/Permissions">per-user authentication</a> on supported connections, so each source applies that user’s own permissions</li><li>Assign roles and permissions in one place, and choose which tools a team or agent sees with <a href="https://docs.cloud.cdata.com/en/Toolkits">Toolkits</a> (Growth tier and above)</li><li>Redact or block sensitive data with <a href="https://docs.cloud.cdata.com/en/Data-Security">data security rules</a></li><li>Review who made each call-in <a href="https://docs.cloud.cdata.com/en/Logs">request-level logs</a>, and forward audit events to your <strong>security information and event management (SIEM)</strong> system on the Business tier</li></ul><p>You don’t need to operate a separate server and gateway. To learn more about Connect AI Gateway, read <a href="https://www.cdata.com/blog/choose-best-enterprise-ai-gateway">how to choose an enterprise AI gateway</a> and the <a href="https://docs.cloud.cdata.com/en/MCP-Gateway">MCP Gateway documentation</a>.</p><h3>Hybrid deployment patterns for optimal balance</h3><p>A hybrid MCP architecture combines direct server connections with gateway-managed access. Organizations can place sensitive or shared tools behind a gateway and keep direct access to selected servers for specialized workflows.</p><p>Common deployment patterns include:</p><ul><li><strong>Team-specific virtual MCP servers</strong>: A gateway gives each team its own view of the available tools while backend servers stay independently managed</li><li><strong>Selective gateway routing</strong>: Regulated or high-risk tools pass through the gateway for policy enforcement and auditing, while approved low-risk tools connect directly</li><li><strong>Phased adoption</strong>: Organizations move existing MCP servers behind a gateway gradually, starting with sensitive workloads and shared tools</li></ul><p>Direct connections in a hybrid design must still enforce authentication, authorization, and logging</p><h3>Run MCP servers behind one governed layer with CData Connect AI Gateway</h3><p>A gateway gives you one place to set access policy and review activity across servers. <a href="https://www.cdata.com/ai/">CData Connect AI Gateway</a> adds governed access to hundreds of enterprise data sources through one endpoint.</p><p><a href="https://www.cdata.com/ai/signup/">Start your free trial today.</a></p><h3>Frequently asked questions</h3><p>What impacts latency when using an MCP gateway?</p><p>The extra routing hop and each control you enable, such as authentication, policy checks, and logging. Measure request time with and without the gateway on your own workloads.</p><p>How does an MCP gateway improve token cost efficiency?</p><p>Many gateways can send each agent a shorter, scoped tool list, which reduces the tokens spent on tool descriptions. The effect grows with the number of servers and varies by product.</p><p>What security advantages does an MCP gateway provide?</p><p>It centralizes authentication, authorization, and audit trails, and it can apply role-based access control and policy enforcement in one place. It’s also a component you need to secure and monitor.</p><p>How do MCP gateways support multi-team governance?</p><p>They centralize access management, policy, and monitoring, so each team gets its own scoped access to data and tools under one set of audit records.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/mcp-gateway-vs-mcpserver-2026"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d6a54e86e4cb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/mcp-gateway-vs-mcp-server-performance-scale-and-control-d6a54e86e4cb">MCP Gateway vs MCP Server: Performance, Scale, and Control</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ETL Tool Performance Benchmarks: How to Evaluate in 2026]]></title>
            <link>https://medium.com/cdata-software/etl-tool-performance-benchmarks-how-to-evaluate-in-2026-98c26f7c0fef?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/98c26f7c0fef</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[data-integration]]></category>
            <category><![CDATA[etl]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[data-architecture]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 15:12:05 GMT</pubDate>
            <atom:updated>2026-11-05T15:12:05.379Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*4kmZTVD8tlu8bI0L" /></figure><p>If you ask a vendor what dataset their benchmark slide is built on, their confidence usually drops fast. Most published numbers come from a lab setup that looks nothing like your stack: a different source system, a different transformation, often hardware you’ll never really run.</p><p>If you lead data architecture at a mid-size or large enterprise, it’s likely that you’ve sat through this pitch before. You need a method you can run yourself, one that holds up when InfoSec starts asking questions and still works the next time a vendor quietly changes its pricing page.</p><h3>What is an ETL performance benchmark?</h3><p>An <a href="https://www.cdata.com/blog/what-is-etl">ETL (Extract, Transform, Load)</a> performance benchmark is a controlled, repeatable test: the same workload, run against each candidate tool, so results line up next to each other instead of past each other.</p><p>No single industry standard covers ETL performance. The <strong>Transaction Processing Performance Council</strong> released <strong>TPC-DI</strong> to fill that void, the first formal benchmark built for data integration workloads. It models a brokerage firm’s data flow and separates timed work (extraction, transformation, loading) from untimed setup (data generation, teardown).</p><p>An ETL performance benchmark is only meaningful if every tool runs against the same dataset, transformations, environment, and success criteria.</p><p>Vendor-published numbers are hard to trust because architectures, connectors, and deployment models differ so much between tools. You need to build a method that you can run instead.</p><h3>Core metrics that matter: throughput, latency, and scalability</h3><p><strong>Throughput</strong> is how much data a pipeline moves per unit of time, usually <em>rows per second</em> or <em>megabytes per second</em>. <strong>Latency</strong> is the time from extraction to availability at the destination, split into event-time latency (how stale the data is) and processing-time latency (how long the pipeline takes).</p><p>The two trade off: research on IoT analytics pipelines found latency from source to destination deteriorates as throughput rises.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G917NO1ni9uRsT6oIqWAwQ.png" /></figure><p>A modern enterprise target runs north of <em>50,000 rows per second</em>, with latency under roughly <em>10 seconds per batch</em>. Treat these as reference points as your own freshness requirements set the real bar.</p><h3>Before you benchmark: prerequisites and evaluation criteria</h3><p>Four things need to be true first: a workload matching your production mix, a dataset sized and skewed like your real data, matched hardware and configuration across every tool, and identical success criteria scored the same way for each.</p><p>Test more than one processing mode where supported; <strong>batch</strong>, <strong>incremental</strong>, and <strong>streaming</strong> performance can differ sharply inside one product, which is why TPC-DI runs a historical load plus two incremental loads. Capture a baseline first, and build in correctness checks from day one, hash checks, data model validation, and record-count reconciliation, the way TPC-DI’s audit checks do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2mVbxnoox4JLLqznAvLOLw.png" /></figure><p>If you want a parallel checklist for the selection side of this problem, our <a href="https://www.cdata.com/blog/best-hybrid-etl-tool-selection-guide-2026">2026 guide to picking a hybrid ETL tool</a> covers similar ground from a different angle.</p><h3>Step 1: Define your benchmark scope and success criteria</h3><p>Write success criteria before any test script: target throughput, maximum p95 latency, a cost ceiling per terabyte. Score every tool pass or fail against those numbers.</p><p>Decide which stages count toward the clock; TPC-DI only times ingestion, transformation, and loading. Use the same rule, or your fastest tool might just be the one that cut scope. Tie targets to real freshness needs: near-real-time ingestion means little if the pipeline downstream only runs a few times a day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qYCvW2buFrU9mjtwNEoLqQ.png" /></figure><p>Document every scope decision. A report only holds up if it states dataset, transformations, environment, runtime, throughput, cost, and validation method clearly enough to rerun.</p><h3>Step 2: Build a representative test dataset and cloud ETL environment</h3><p>Size your dataset to expose bottlenecks, then size it again for growth. TPC-style benchmarks use a <strong>Scale Factor</strong>, a multiplier that produces proportionally larger datasets so one workload runs at 100 GB, 500 GB, 1 TB, and beyond.</p><p><strong>Quick definition: Scale Factor.</strong> A multiplier used by benchmark generators to produce proportionally larger datasets for the same workload.</p><p>TPC-DI results show the historical load consistently takes about 80% of total elapsed time regardless of data size, with the two incremental loads splitting the rest. Test at more than one scale factor to see whether your candidate scales linearly or degrades.</p><p>Hold compute configuration constant across every tool, normalizing performance to price point with enterprise-equivalent tiers, so a budget tier on one platform never gets compared against a premium tier on another. Generate load with mature tooling rather than hand-rolled data, and simulate production noise, compliance deletes, backfills, key rotation, since background activity changes measured throughput.</p><h3>Step 3: Measure throughput and latency under real-world load</h3><p>Instrument <strong>extraction</strong>, <strong>transformation</strong>, and <strong>loading</strong> separately so a spike points to a stage. Measure at the driver, per stage, and from source to destination, the three-point approach stream-processing research recommends for isolating where a slowdown starts.</p><p>Wire up Prometheus or Grafana (or whatever your cloud already gives you) and watch resource use against execution time while the job runs. That’s where you see a bottleneck form.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Pftm5B7785hOoZVqkeXgQ.png" /></figure><h3>Step 4: Test scalability and concurrency for enterprise workloads</h3><p>Build concurrency into the test: run incremental loads alongside background reads and writes, the way production behaves, and report throughput under that load.</p><p><strong>Quick definition: sustainable throughput.</strong> The highest data rate a pipeline can process continuously without latency growing unbounded or a backlog forming.</p><p>Latency climbs as throughput rises, and checkpoint and recovery settings trade service quality against recovery speed. Record where each tool’s latency curve breaks as concurrency increases; that breakpoint matters more than the peak number on a slide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HOyt6CUJ0V2bGLU9iJfjrw.png" /></figure><p><em>Illustrative figures, not vendor-reported results.</em></p><h3>Step 5: Compare connector depth across Snowflake, Databricks, and AWS</h3><p>Connector depth covers driver efficiency, live connectivity, and how much transformation logic a connector pushes down before data crosses the network.</p><p><strong>Quick definition: query pushdown.</strong> A technique that translates transformation logic into SQL executed at the source or target database, avoiding costly data movement.</p><p>Pushdown pushes filters, joins, and aggregations into the data source instead of pulling everything across the network into the ETL engine. Skip it, and the ETL server absorbs the full load, which bottlenecks fast on large datasets. Snowflake’s pushdown shifts real processing weight onto Snowflake itself. AWS Glue’s cuts scanned data, but unevenly; its DynamoDB reader skips pushdown predicates entirely.</p><p>CData’s standards-based JDBC and ODBC drivers give live connectivity to Snowflake, Databricks, and hundreds of other sources, with pushdown bounded by what the source and driver can support, a solid reference point for what connector depth should mean here.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DXLKrLN6YeIzTDJj6IdXeg.png" /></figure><h3>Step 6: Evaluate price-performance and total cost of ownership at scale</h3><p>ETL often eats the biggest chunk of cloud data spend, over half of it by one estimate. That’s reason enough to rank tools on price-performance when you’re the one justifying the budget.</p><p>Measure runtime and cost on the same workload, then compute cost-per-gigabyte and normalize each platform to its price point. Newer Snowflake compute generations run faster but cost more, enough to flip a ranking that looked settled on speed alone.</p><p>AWS Glue gives a concrete anchor: <em>$0.44 per DPU-hour</em>, billed per second with a one-minute minimum, where one DPU is 4 vCPUs and 16 GB of memory. Six DPUs for 42 minutes runs about <em>$1.85 in compute</em>, a number that scales cleanly to a production estimate. One catch: consumption-based pricing swings with usage, so budget off a load-scaled projection, not the price of one test run.</p><p>Row-based and DPU-based pricing carry the same forecasting problem; our look at <a href="https://www.cdata.com/blog/understanding-usage-based-pricing-for-etl-software">usage-based ETL pricing</a> uncovers how that plays out across a few common models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vrdPoXD6OXZvxAIyIfgQEg.png" /></figure><p><em>Rate from AWS Glue’s published pricing; cost-per-TB is illustrative.</em></p><h3>Interpreting your results: a side-by-side comparison framework</h3><p>Put every result in one scorecard, dataset, environment, runtime, throughput, cost, scaling behavior, and validation method all in the same row. If you leave one out, nobody can check your work. Weigh the metrics by what your team cares about: cost and speed rarely rank the same tool first. And add a correctness gate, a column that confirms every candidate passed reconciliation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WrFUMu9hTYLvwNzz02rqHg.png" /></figure><p><em>Illustrative row; populate with your own results.</em></p><p>Rank candidates on price-performance and reliability. Throughput alone picks the wrong tool.</p><h3>Practical tips and troubleshooting for reliable benchmarks</h3><ul><li>Add automated performance gates to CI/CD that fail a build on a meaningful execution-time spike, and track history to catch drift early</li><li>Correlate CPU, memory, disk I/O, and network per run against execution time to find the resource-bound stage, comparing variants rather than absolute numbers</li><li>Rule out hardware first: limited I/O, insufficient RAM, and an under-provisioned CPU cap throughput regardless of the ETL logic</li><li>Average metrics over a window and measure after several incremental loads, not a cold start, so results reflect stable-state performance</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M8WgXK4eZI37EcPXEUQkkQ.png" /></figure><h3>Common ETL benchmarking mistakes to avoid</h3><ul><li><strong>Comparing tools on unequal footing.</strong> Different datasets, transformations, or default write modes, append-only versus upsert, can skew results on their own</li><li><strong>Measuring only total runtime.</strong> Without per-stage insight, teams optimize blind</li><li><strong>Ignoring concurrency and background processes.</strong> Skipping concurrent reads, writes, and incremental loads overstates real-world throughput</li><li><strong>Skipping correctness validation.</strong> A performance win means nothing if output fails hash checks or record reconciliation</li><li><strong>Trusting a single vendor’s benchmark at face value.</strong> No universal standard exists; reproduce every result in your own environment first</li></ul><h3>Build a defensible ETL evaluation with CData Sync</h3><p>Based on an <a href="https://www.cdata.com/lp/sync-replication-report/">internal benchmarking report</a>, <a href="https://www.cdata.com/sync/">CData Sync</a> replicated a 360 GB, 1-billion-row SQL Server table into Databricks in 2 hours 41 minutes, down from 11 hours 6 minutes on the prior version, same infrastructure, same workload.</p><p>Run this framework against your own shortlist, including a configuration built on CData Sync’s standards-based connectivity to Snowflake, Databricks, and hundreds of other sources. Do not take any vendor’s published numbers, including this one, without rerunning the test yourself.</p><p><a href="https://www.cdata.com/sync/trial/">Start a free 30-day trial of CData Sync</a><strong>.</strong></p><h3>Frequently asked questions</h3><p>What metrics should I include in an ETL tool performance benchmark?</p><p>Throughput, latency, resource utilization, error rate, cost per run, and scaling behavior under growth.</p><p>How do I set up a fair ETL benchmark test so the results are comparable?</p><p>Same dataset, transformations, environment, and hardware tier for every tool, scored against identical success criteria.</p><p>What dataset size should I use when benchmarking ETL pipeline performance?</p><p>Production-scale volumes, tested at more than one size, 100 GB, 500 GB, 1 TB and beyond, to see if performance holds or degrades.</p><p>Should I benchmark batch, incremental, and streaming ETL workflows separately?</p><p>Yes. Performance can differ sharply by mode inside the same tool, so test each mode you’ll run.</p><p>How do I validate that my ETL benchmark results are accurate and not just fast?</p><p>Run hash checks, data model validation, and record-count reconciliation on every test. Speed means nothing if rows go missing.</p><p>Does a faster ETL tool always mean lower total cost, and how do I evaluate cost-performance?</p><p>No. Faster tools can cost more per run. Compare cost-per-terabyte normalized to price point, not raw speed.</p><p>Why is it so hard to find standardized ETL tool performance benchmarks, and what should I do instead?</p><p>No universal standard exists. Run the test yourself, same dataset and environment, across every tool on your shortlist.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/etl-tool-performance-benchmarks-2026"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98c26f7c0fef" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/etl-tool-performance-benchmarks-how-to-evaluate-in-2026-98c26f7c0fef">ETL Tool Performance Benchmarks: How to Evaluate in 2026</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Copilot Agent Data Access: How to Add Enterprise Data (2026 Guide)]]></title>
            <link>https://medium.com/cdata-software/copilot-agent-data-access-how-to-add-enterprise-data-2026-guide-0e78d088b09f?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/0e78d088b09f</guid>
            <category><![CDATA[microsoft-copilot]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 14:37:53 GMT</pubDate>
            <atom:updated>2026-11-05T14:37:53.456Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*WR-_x4HwpPB7motR" /></figure><p>To add enterprise data to a Microsoft Copilot agent safely, connect the agent to a <a href="https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro">Model Context Protocol (MCP)</a> server that reaches your systems of record, then carry each user’s identity and permissions through to every source.</p><p>Skip that second step and the agent either answers from general model knowledge or reads data through a shared service account that can see more than any one user should.</p><p>The <a href="https://www.cdata.com/company/press/cdata-microsoft-enterprise-ai-agents-mcp-integration/">Microsoft and CData partnership</a> addresses this. It brings Connect AI’s MCP connectivity into Copilot Studio and Microsoft Agent 365, so agents can read, write, and act on real-time data in systems like Salesforce, SAP, and NetSuite.</p><p>Your security team needs to know who can see what and how access is audited. Your engineering team needs a build that holds up in production. This guide covers setup, per-user permissions, governance, and observability for both.</p><h3>What copilot agent data access means (and why native connections fall short)</h3><p>With Copilot agent data access, an agent built in Copilot Studio pulls live business data when a user asks a question. Microsoft states that Copilot <a href="https://learn.microsoft.com/microsoft-365-copilot/microsoft-365-copilot-privacy">only surfaces data</a> a user has at least view permission to access in Microsoft Graph. Your Salesforce pipeline, SAP orders, and Workday records sit outside Microsoft Graph, so that check doesn’t apply to them.</p><p>To reach those systems, the agent needs a tool, and MCP has been the standard way to add one since Microsoft made it <a href="https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/model-context-protocol-mcp-is-now-generally-available-in-microsoft-copilot-studio/">generally available</a> in Copilot Studio in May 2025. Most teams start by connecting <a href="https://www.cdata.com/blog/overcome-data-silos-with-mcp-gateways">one server per source</a>, which holds up for a pilot with one or two systems.</p><p>At dozens of sources, your team maintains separate tools, credentials, and policies for each one, and an audit request means stitching logs together from every server. CData’s <a href="https://www.cdata.com/ai/integrations/microsoft/">Copilot integration page</a> also notes that without a managed layer, Copilot filters, joins, and groups datasets in code, which in some cases uses up to 10X more tokens.</p><p>A centralized gateway replaces those servers with a single connection point. CData Connect AI works as that governed gateway, sitting between Copilot and hundreds of enterprise systems, with permissions and logging handled in one place.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-U878MNDzcnExOvfyeydBQ.png" /></figure><p>The result is one connection to secure, monitor, and audit, however many sources sit behind it.</p><h3>Prerequisites: Copilot Studio capabilities, licensing, and governance requirements</h3><p>Copilot Studio is the copilot agent builder for Microsoft 365. You create an agent, write its instructions, add knowledge, and attach tools before you publish it. Makers typically get access through a free Copilot Studio user license or a Microsoft 365 Copilot license. Microsoft notes that the <a href="https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing">free user license</a> requires a tenant prepaid Copilot Credit pack first.</p><p>Runtime usage is billed in Copilot Credits. For Microsoft 365 Copilot licensed users, classic answers, generative answers, and Microsoft Graph grounding in Copilot Chat, Teams, or SharePoint don’t count against the meter. Plan credit capacity for any other usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6ofG1wFhQNEN8EOV7pmfrg.png" /></figure><p>For a faster security review, settle the governance decisions before anyone builds. Decide which sources need per-user access, whether admins will block shared maker credentials, and who approves new MCP servers. Agents built after those calls are made don’t need to be reworked for security review later.</p><h3>How to connect enterprise data to a Copilot agent via MCP (step by step)</h3><p>Each step below closes one of the gaps that per-source integrations leave open. By the end, every request from the agent carries the signed-in user’s identity to the source.</p><h3>Step 1: Set up your AI gateway connection in CData Connect AI</h3><p>Sign in to <a href="https://cloud.cdata.com/">Connect AI</a> or start a trial. If your organization uses single sign-on (SSO), you can have CData enable it for your account so people sign in to Connect AI with their corporate identity.</p><p>Every source you add sits behind one <a href="https://docs.cloud.cdata.com/en/API/MCP">MCP endpoint</a>, <a href="https://mcp.cloud.cdata.com/mcp,">https://mcp.cloud.cdata.com/mcp,</a> which you enter as the server URL in clients that ask for one. Connect AI supports OAuth and Basic authentication, and Basic uses your Connect AI user ID and a personal access token (PAT). A PAT ties every request to one account, so keep it for testing and use OAuth for production agents.</p><h3>Step 2: Add and configure enterprise sources like Salesforce and SAP</h3><ol><li>In Connect AI, open <strong>Sources</strong> and select <strong>Add Connection</strong>.</li><li>Add each system the agent needs, such as Salesforce CRM, SAP ERP, or Workday.</li><li>Where the source supports it, switch the connection to <a href="https://docs.cloud.cdata.com/en/Permissions">per-user authentication</a> so each user signs in to the source with their own account.</li><li>On the <strong>Permissions</strong> tab, set Select, Insert, Update, Delete, and Execute rights for each user.</li></ol><p>Connections use shared credentials by default. That’s reasonable for read-only reference data with no user-specific permissions. For CRM, ERP, and HR data, per-user authentication is what keeps the source system’s role-based access control (RBAC) in effect when a user asks the agent a question.</p><h3>Step 3: Attach the MCP toolset to your agent in the Copilot agent builder</h3><p>Open the agent in Copilot Studio, go to the <strong>Tools</strong> page, and select <strong>Add a tool</strong>. From there, you can add the certified <a href="https://learn.microsoft.com/en-us/connectors/cdataconnectai/">CData Connect AI connector</a>, which has users sign in with their Connect AI account. You can also use the MCP onboarding wizard with the endpoint from Step 1.</p><p>In the tool’s settings, set <strong>Credentials to use</strong> to <strong>End user credentials</strong>. Each user then signs in the first time the agent calls the tool. The other option, maker-provided credentials, runs every request as the person who built the agent. Admins can <a href="https://learn.microsoft.com/en-us/power-platform/release-plan/2026wave1/microsoft-copilot-studio/block-use-maker-provided-credentials-authentication">block maker credentials</a> across an environment from the Power Platform admin center.</p><p>For a full walkthrough, follow the <a href="https://www.cdata.com/kb/articles/cloud-copilot-teams.rst">Copilot Studio and Teams</a> guide. If your source is a REST API, the <a href="https://www.cdata.com/kb/tech/api-cloud-copilot.rst">API data guide</a> covers that setup.</p><h3>Security and governance: RBAC, OAuth/SSO passthrough, and Agent 365 controls</h3><p>A security lead will usually start by asking whether a user can see more through the agent than they could by logging in to the source directly. That depends on four links in the identity chain.</p><ol><li>The user signs in to Microsoft 365 with Microsoft Entra ID.</li><li>End-user credentials send the request to Connect AI as that user.</li><li>Connect AI applies its per-user permissions, with SSO through your identity provider.</li><li>Per-user authentication sends the query to the source, where its RBAC applies unchanged.</li></ol><p>Together, these links are OAuth/SSO passthrough. If any link falls back to a shared identity, the agent can return data the user can’t see on their own. With all four in place, the source system decides what comes back, the same way it does when the user logs in directly.</p><p>Because every request passes through one gateway, your security lead reviews one set of identity settings and one log, instead of auditing a separate integration for each source.</p><p><a href="https://learn.microsoft.com/en-us/microsoft-agent-365/overview">Microsoft Agent 365</a>, generally available since May 1, 2026, governs the agents. Admins get one registry to observe and manage agents, and they can <a href="https://learn.microsoft.com/en-us/microsoft-365/admin/manage/manage-tools-for-agent">approve or reject</a> custom MCP servers in the Microsoft 365 admin center.</p><p>Connect AI governs the data access underneath with its query log, which records the timestamp, user, query text, and status for every query. A separate audit log tracks administrative changes such as new users and permission edits. The <a href="https://www.cdata.com/blog/enterprise-mcp-security-playbook">MCP security playbook</a> covers how to route these logs into your monitoring tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lz9szF-HEZSVdUpdXEDfDg.png" /></figure><h3>Troubleshooting: why your Copilot agent still gives generic answers</h3><p>A generic answer usually means the agent never reached your systems and fell back on general model knowledge. A few causes account for most cases.</p><ul><li><strong>Missing tool:</strong> Add the Connect AI tool to the agent and turn on generative orchestration, which MCP tools require.</li><li><strong>No sign-in:</strong> With end-user credentials, each user has to sign in to the connection before the agent can call it.</li><li><strong>No source access:</strong> If the user has no rights in the source, an empty result is correct and shows the permission chain is working.</li><li><strong>Missing context:</strong> The agent can reach a source and still miss the right field. Connect AI’s per-source MCP instructions describe each system’s structure and business definitions, so the agent finds the right data.</li></ul><p>Most of these take a few minutes to rule out in Copilot Studio and the Connect AI console. Run this check before you escalate.</p><ol><li>Look for the Connect AI tool on the agent’s <strong>Tools</strong> page.</li><li>Check that generative orchestration is turned on.</li><li>Make sure the user has signed in to the connection.</li><li>Verify the user has access to both the source and the Connect AI connection.</li><li>Open the activity map in Copilot Studio to see which MCP tool the agent called.</li></ol><h3>Bring enterprise data in your Copilot agent with CData Connect AI</h3><p>With one governed AI gateway connection in place of dozens of point-to-point integrations, your security team approves a single layer that inherits the permissions your source systems already enforce.</p><p>To build your first Copilot agent, follow the <a href="https://www.cdata.com/kb/articles/connect-ai-copilot-start-guide.rst">Connect AI quickstart</a>, which sets up a Copilot Studio agent on sample data in a few minutes.</p><p>When you’re ready for Salesforce, SAP, or Workday, see the <a href="https://www.cdata.com/ai/integrations/microsoft/">Copilot integration</a> page and start a 14-day free trial of CData Connect AI.</p><h3>Frequently asked questions</h3><p>What types of enterprise data can a Copilot agent access?</p><p>Inside Microsoft 365, an agent can use Microsoft Graph content such as files, mail, and Teams messages. Through CData Connect AI, it can also query hundreds of outside systems, including Salesforce, SAP, NetSuite, Snowflake, and Workday.</p><p>Does a Copilot agent access data that the user isn’t already authorized to see?</p><p>For Microsoft Graph content, no. For external sources, it depends on the tool’s credentials. With end-user credentials in Copilot Studio and per-user authentication in Connect AI, the agent returns only what that user can already see.</p><p>Is it safe to use confidential company data with a Copilot agent?</p><p>It can be, once identity and logging are set up correctly. Microsoft states that prompts, responses, and Microsoft Graph data aren’t used to train Copilot’s foundation models. For outside systems, per-user authentication keeps source permissions in force, and Connect AI logs every query.</p><p>What do I need before I can connect a Copilot agent to enterprise data sources like Salesforce or SAP?</p><p>You need Copilot Studio access, Copilot Credits or Microsoft 365 Copilot licenses, and Microsoft Entra identity for sign-in. You also need a CData Connect AI account with at least one source connected. Agent 365 is optional.</p><p>Why is my Copilot agent giving generic answers instead of pulling from my company’s live data?</p><p>Usually the agent never reached your systems. Check that the Connect AI tool is attached, generative orchestration is on, and the user has signed in to the connection. The activity map in Copilot Studio shows which tool the agent called.</p><p>How do admins control which data sources and users a Copilot agent can access?</p><p>In the Power Platform admin center, admins can block maker-provided credentials so every tool runs as the signed-in user. In Agent 365, they approve or reject MCP servers. In Connect AI, they set per-user authentication and operation-level permissions on each connection.</p><p>What’s the difference between a Copilot agent and Copilot Studio when it comes to enterprise data access?</p><p>A Copilot agent is the assistant your users talk to in Microsoft 365 Copilot, Teams, or SharePoint. Copilot Studio is where you build it and attach tools such as MCP servers, which is how it reaches data outside Microsoft 365.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/copilot-agent-data-access"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e78d088b09f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/copilot-agent-data-access-how-to-add-enterprise-data-2026-guide-0e78d088b09f">Copilot Agent Data Access: How to Add Enterprise Data (2026 Guide)</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SAP to Agent Bricks: 2026 Guide to Connecting SAP ERP]]></title>
            <link>https://medium.com/cdata-software/sap-to-agent-bricks-2026-guide-to-connecting-sap-erp-d135fa216224?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/d135fa216224</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[sap]]></category>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 14:23:40 GMT</pubDate>
            <atom:updated>2026-11-05T14:23:40.747Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*Nq0XseBOjinEW5Jl" /></figure><p>Connecting SAP enterprise resource planning (ERP) data to Databricks <strong>Agent Bricks</strong> is a governed-access problem, and CData Connect AI solves it as an AI gateway that lets an agent fetch live SAP data using inherited user permissions. Replicated SAP data needs its own pipelines, sync schedules, and permission models. To connect SAP ERP data to an Agent Bricks agent, add the Connect AI <strong>Model Context Protocol (MCP)</strong> server from Databricks Marketplace, authenticate SAP, and build the agent.</p><h3>What is Databricks Agent Bricks?</h3><p>Agent Bricks is Databricks’ <a href="https://www.databricks.com/blog/agent-bricks-governed-enterprise-agent-platform">enterprise agent platform</a> for building, deploying, and governing agents that operate on business data, unifying model access, execution, governance, and context in one system. The harder work in enterprise agents is making them operate on real business data, under real permissions, with real consequences. <strong>Document Intelligence</strong> and <strong>Custom Agents</strong> are generally available, and production agents run across financial services, retail, healthcare, and technology.</p><h3>What SAP ERP connectivity via MCP means for Agent Bricks</h3><p>MCP is an open-source standard that connects AI agents to tools, resources, prompts, and other contextual information. Databricks supports <a href="https://docs.databricks.com/aws/generative-ai/mcp">external MCP servers</a>, which run outside Databricks and install from Databricks Marketplace or connect through <strong>Unity Catalog</strong> connections. Adding the Connect AI MCP server therefore gives the agent live access to SAP through a standardized interface, with no extract, transform, and load ( <a href="https://www.cdata.com/blog/sap-to-databricks-etl-made-easy">ETL</a>) pipeline in between. The AI gateway and the MCP server are the same component seen from two sides: Databricks registers and governs the server, and Connect AI enforces source permissions behind it. SAP’s <a href="https://www.cdata.com/blog/sap-api-policy-ai-strategy">API policy</a> also shapes how enterprises plan AI connectivity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z9WSwNDhbc9rngfMZShc3w.png" /></figure><h3>Why connect SAP ERP to Agent Bricks with CData Connect AI</h3><h3>A Databricks-validated MCP Marketplace launch partner</h3><p>CData <a href="https://www.cdata.com/company/press/cdata-featured-mcp-launch-partner-databricks-marketplace/">partnered with Databricks</a> to offer Connect AI, an AI gateway, on Databricks Marketplace. Connect AI enables Agent Bricks to connect directly to hundreds of enterprise data systems through the MCP protocol, including SAP, NetSuite, Salesforce, and ServiceNow. Governed write-back extends that reach from context to action.</p><h3>Production-ready connectivity in minutes</h3><p>CData hosts and manages Connect AI, so teams configure a data source in minutes. Point-and-click setup and a free tier keep the barrier to a pilot low.</p><h3>Prerequisites and SAP ERP module compatibility</h3><h3>Prerequisites to verify before starting</h3><ul><li><strong>Workspace with Unity Catalog enabled:</strong> Unity Catalog governs all data and AI assets, including MCP services, and Databricks <a href="https://docs.databricks.com/en/data-governance/unity-catalog/">enables it automatically</a> for workspaces created after November 8, 2023. MCP services are securable objects ( catalog.schema.mcp_service), so the Connect AI MCP server receives the same governance as any other asset. External MCP servers register as MCP services once the account enables <strong>Unity AI Gateway</strong> (Beta) and <strong>Managed MCP Servers</strong> (preview).</li><li><strong>Support for Open Authorization (OAuth) 2.0 in SAP:</strong> Confirm the SAP source supports OAuth 2.0 before starting.</li><li><strong>Open Data Protocol (OData) services:</strong> SAP ERP data reaches external consumers through OData services on <strong>Application Server ABAP (AS ABAP)</strong> and <strong>SAP Gateway</strong>. Confirm each required module exposes an OData service.</li></ul><h3>Supported SAP modules and data sources</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bi-tW9_HVejSPHjbR2N-6g.png" /></figure><h3>How to connect SAP ERP data to Agent Bricks (step-by-step)</h3><h3>Step 1: Authenticate the SAP ERP source in CData Connect AI</h3><p>Connect AI uses point-and-click setup. Add SAP Netweaver Gateway as a connection in the Connect AI console and authenticate to the SAP source.</p><h3>Step 2: Add the Connect AI MCP server from the Databricks MCP Marketplace</h3><p>Create an <a href="https://www.cdata.com/kb/tech/sapgateway-cloud-databricks-agentbricks.rst">OAuth app</a> in Connect AI, then add a Unity Catalog connection to the Connect AI MCP server. Databricks Marketplace lists the Connect AI MCP server.</p><h3>Step 3: Configure OAuth authentication and user permissions</h3><p>Databricks Agent Bricks authenticates to the Connect AI MCP server with OAuth machine-to-machine authentication. The OAuth app runs under the Connect AI user account selected during setup, and Connect AI <a href="https://www.cdata.com/ai/it-approval/">applies the permissions</a> that user already holds in the connected source. SAP also supports user-delegated access through the <a href="https://help.sap.com/saphelp_em92/helpdata/en/12/41087770d9441682e3e02958997846/content.htm">OAuth 2.0 Security Assertion Markup Language (SAML) Bearer Assertion flow</a>:</p><ul><li>The identity provider issues a SAML assertion.</li><li>The calling application exchanges the assertion for an OAuth access token.</li><li>The application uses the token to reach SAP OData resources on behalf of the user, within the granted OAuth 2.0 scopes.</li></ul><h3>Step 4: Build the Agent Bricks agent on live SAP data</h3><p>In Agent Bricks, define the task and connect the data sources. Agent Bricks auto-generates evaluation sets and tunes the agent against them. The agent template connects to MCP servers, so the Connect AI MCP server appears as a set of tools for fetching and acting on SAP objects.</p><h3>Configuring Unity Catalog governance and least-privilege controls</h3><h3>Enforcing existing SAP permissions at runtime</h3><p>Unity Catalog is the unified governance layer for data and AI in Databricks, providing centralized access control, lineage, auditing, and data discovery across workspaces. It operates beneath every data and AI interaction automatically, and it governs the Connect AI MCP server as an MCP service. Connect AI enforces existing source-system permissions at the moment of request, with no data replication, so no parallel permission model needs maintenance.</p><h3>Audit trails and compliance for production rollout</h3><p>Unity Catalog records every invocation of the MCP service in its audit log system table, and Connect AI <a href="https://docs.cloud.cdata.com/en/Logs">logs each request to SAP</a> with the Connect AI user, timestamp, and executed request. Together, the two logs give a Databricks-side and source-side audit trail.</p><h3>Real-world SAP + Agent Bricks use cases</h3><h3>Finance reconciliation agent</h3><p>A finance reconciliation agent reads live SAP finance data, such as general ledger accounts, and matches open items against it. Write-back lets it act on records once it finds a discrepancy.</p><h3>Supply chain order-processing agent</h3><p>Databricks reports that teams build agents orchestrating supply chain, procurement, and R&amp;D workflows. Databricks says zero-copy access to <strong>SAP Business Data Cloud</strong> data will enable <a href="https://www.databricks.com/blog/announcing-general-availability-sap-business-data-cloud-connect-databricks">autonomous decision-making</a> in areas such as supply chain optimization and demand forecasting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GjJVnotjjW7D5mfydktnuw.png" /></figure><h3>Troubleshooting and best practices</h3><ul><li><strong>Confirm the OAuth app user:</strong> Select a Connect AI user account that has access to the SAP connection.</li><li><strong>Enable audit logging before launch:</strong> Confirm audit events appear in Unity Catalog audit logs before launch.</li></ul><h3>Start building SAP-connected Agent Bricks agents with CData Connect AI</h3><p>CData Connect AI is immediately available in Databricks Marketplace. Configure a data source for Agent Bricks in minutes and start building.</p><p>As a validated MCP launch partner, CData brings that connectivity to SAP ERP data under existing source permissions. <a href="https://www.cdata.com/ai/signup/">Start a free trial today</a>.</p><h3>Frequently asked questions</h3><p>What do I need before connecting SAP ERP data to Databricks Agent Bricks?</p><p>A Unity Catalog-enabled workspace with Unity AI Gateway (Beta) and Managed MCP Servers (preview) enabled, an SAP source that supports OAuth 2.0, and OData services for each required module. Confirm all three first.</p><p>How do I connect SAP to Agent Bricks using the Databricks MCP Marketplace?</p><p>Authenticate SAP in Connect AI, add its MCP server from Databricks Marketplace as a Unity Catalog connection, configure OAuth authentication, and build the agent. Create the OAuth app first.</p><p>Does connecting SAP ERP to Agent Bricks require custom ETL pipelines or coding?</p><p>No ETL pipeline is required, because agents request live SAP data through the MCP server. Setup uses point-and-click configuration.</p><p>How does Agent Bricks access SAP data without moving or copying it?</p><p>Each request travels through the MCP server to SAP at runtime. SAP stays the system of record, so no replica exists.</p><p>How do I enforce SAP user permissions and audit trails in a Databricks Agent Bricks agent?</p><p>The Connect AI OAuth app runs as a designated Connect AI user, and Connect AI applies that user’s existing permissions. Unity Catalog captures user-level audit logs automatically.</p><p>What SAP ERP use cases work best with Databricks Agent Bricks in 2026?</p><p>Finance reconciliation, supply chain order processing, and procurement monitoring map directly to FI, SD, and MM data. Write-back adds governed actions.</p><p>Why is my Agent Bricks agent returning inaccurate or hallucinated SAP data, and how do I fix it?</p><p>Start with scope: the agent sees only what the Connect AI user’s permissions and the OData services expose. Then tighten the task description and review answers in evaluation.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/sap-agent-bricks-connect-sap-erp-2026"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d135fa216224" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/sap-to-agent-bricks-2026-guide-to-connecting-sap-erp-d135fa216224">SAP to Agent Bricks: 2026 Guide to Connecting SAP ERP</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Enterprise MCP Gateway: Securely Run MCP Behind the Firewall]]></title>
            <link>https://medium.com/cdata-software/enterprise-mcp-gateway-securely-run-mcp-behind-the-firewall-e2f75d508477?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/e2f75d508477</guid>
            <category><![CDATA[security]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 14:05:44 GMT</pubDate>
            <atom:updated>2026-11-05T14:05:44.201Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/550/0*qLb_PninbOBKa-Au" /></figure><p><a href="https://www.cdata.com/blog/what-is-mcp">Model Context Protocol (MCP)</a> gives AI assistants a standard way to reach the systems that hold enterprise data. It doesn’t define who’s allowed to see what, which user triggered a given call, or how a security team reconstructs that history during an audit. Anthropic introduced MCP as an open standard in November 2024, and left access control, identity mapping, and audit trails for whoever deploys it to work out.</p><p>For an enterprise architect evaluating MCP at scale, that’s the decision to make: <a href="https://www.cdata.com/blog/managed-mcp-vs-self-hosted">build a gateway in-house or adopt one</a> that already enforces governance across every system the organization runs.</p><h3>What is an enterprise MCP gateway?</h3><p>An enterprise MCP gateway is an access layer between AI agents and the MCP servers connecting to your systems. It gives agents <em>one governed entry point</em> for discovering tools and calling them, and handles authentication, routing, policy enforcement, and audit logging in that one place instead of scattering those jobs across every server.</p><blockquote><strong><em>Quick definition:</em></strong><em> An MCP gateway is a routing and security layer that intercepts agent requests before they reach MCP servers, enforcing authentication, authorization, policy, and audit logging on every call.</em></blockquote><p>The gateway also solves what’s often called the <strong>N×M problem</strong>. Connect N agents directly to M tools and you get N×M custom integrations, each with its own auth flow and its own credential store. A gateway collapses that to one governed connection.</p><p><a href="https://www.cdata.com/ai/">CData Connect AI Gateway</a> is built around this model: one governed connection in front of hundreds of source systems, rather than a custom integration for every agent-to-tool pairing.</p><h3>How an enterprise MCP gateway differs from a standard MCP server</h3><p>An MCP server exposes tools and connects to a specific system, whether that’s a CRM, a database, or an internal API. The gateway governs access to those servers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0B1kg6BtRh-uHvm1hSgZTw.png" /></figure><h3>Why firewall-contained MCP deployment matters for the enterprise</h3><p>Early MCP deployments ran client and server on the same machine, with little need for authorization. That changes once agents connect to remote servers holding sensitive data. AI agents today don’t just generate text. They pull records from databases, update tickets, send emails, and trigger downstream automation, so once an agent can act, containment matters as much as prompt design.</p><p>The <a href="https://www.cdata.com/blog/mcp-security-enterprise-nsa-advisory">NSA’s May 2026 security guidance on MCP</a> names the specific risk: MCP servers rely on bearer tokens as defined in OAuth 2.1 without specifying lifecycle management for refreshing, revoking, or limiting reuse, and that gap can lead to session replay or unauthorized reuse of a valid session.</p><h3>Credential exposure and shadow IT risk</h3><p>The MCP specification calls out token passthrough as an anti-pattern: a server accepting tokens that weren’t issued for it, then forwarding them downstream. That circumvents rate limiting and breaks accountability, because the resource server’s logs show a different identity than the one making the call. A governed gateway uses one OAuth client to your identity provider instead of a sprawl of ad hoc per-user MCP servers, each with its own credential store nobody else can see.</p><p>Connect AI Gateway replaces that sprawl with a single OAuth client to your identity provider and identity passthrough to every connected system, so each source enforces its own permissions for the requesting user.</p><h3>Audit gaps and compliance blind spots</h3><p>A bare MCP transport gives you no rate limiting and no audit trail beyond whatever the agent happens to log on its own. Passing raw tokens compounds the problem: once a server forwards a token it didn’t validate, the downstream system’s logs record the wrong identity, so a compliance team reviewing access history is working from a record that doesn’t match what happened.</p><h3>Scope creep and over-permissioned access</h3><p>An agent holding a broad database or filesystem credential can be manipulated through prompt injection or a poisoned tool into using that credential for something nobody intended, like pulling a full customer table. A gateway constrains this by presenting each agent only the tools it’s authorized to see. An agent scoped to support tasks never sees database administration or deployment tools, because those tools are never on its list.</p><h3>How MCP works behind the firewall</h3><p>MCP servers form the execution plane, connecting to the systems that hold the data: CRMs, databases, ticketing tools, internal APIs. The gateway is the control plane. It decides which agent sees which tools, under which identity, with what constraints.</p><blockquote><strong><em>Quick definition:</em></strong><em> Identity propagation is the gateway’s ability to carry a real user’s identity across a sequence of tool calls, so every downstream action traces back to the person who triggered it rather than a shared service account.</em></blockquote><p>For transport, the MCP authorization spec expects HTTP-based connections to follow its auth rules, with stdio reserved for local processes and HTTPS for anything crossing a network, private or not.</p><h3>Routing MCP connections inside a private network boundary</h3><p>The gateway presents a unified tool catalog and routes each request to the correct upstream server, so an agent gets a consistent, permission-aware view limited to what it’s authorized to use. Microsoft documents a version of this pattern in Azure: an AI gateway in front of MCP tools gives you one governed entry point to enforce authentication, rate limits, IP restrictions, and audit logging without touching the MCP servers or the agent code behind them. That governance layer in Microsoft Foundry is currently a preview feature, and logging runs through Azure API Management and Azure Monitor rather than a single unified service.</p><p>Connect AI Gateway applies the same routing model without requiring an Azure-specific deployment, so the same policy layer covers non-Microsoft stacks too.</p><h3>Runtime identity propagation vs. connection-time credential snapshots</h3><p>A gateway should carry the user’s identity at call time, not capture a static credential once when the connection opens. That’s what makes firewall trust something the system enforces.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tZ-21wmQ1F3MZlHos4ZF2Q.png" /></figure><p>The MCP spec backs this with a hard rule: servers must not accept tokens that weren’t issued for them and must act as the authoritative party instead of forwarding credentials they can’t validate.</p><p>Connect AI Gateway implements this as passthrough identity: each user’s own credentials pass through to the source system at execution time, so the AI agent inherits that user’s existing permissions instead of running under a shared service account.</p><h3>Security controls every enterprise MCP gateway needs</h3><p>The MCP authorization spec mandates two patterns worth treating as a baseline: per-client consent, meaning a user’s general login doesn’t imply authorization for any specific client, and audience validation, meaning every token must verify it was issued for the server receiving it, per <strong>OAuth 2.1 section 5.2 </strong>and <strong>RFC 8707</strong>. Beyond that baseline, a working gateway needs authentication, role-based authorization, rate limiting, IP restrictions, and audit logging in one place.</p><p>Connect AI Gateway bundles all of this by default rather than requiring separate tooling for each control.</p><h3>OAuth/SAML identity passthrough and SSO integration</h3><p>MCP authorization runs on OAuth 2.1, with the MCP server acting as a resource server and a separate authorization server issuing tokens under PKCE. An enterprise gateway connects to your existing identity provider instead of making you stand up a new one, supporting OAuth for token-based auth and SAML for single sign-on. CData Connect AI documents this as <a href="https://www.cdata.com/ai/capabilities/identity-and-access/">passthrough identity with SSO and SAML integration</a>.</p><h3>Workspace isolation and least-privilege enforcement</h3><p>Least privilege means a scoped endpoint exposes only an allow-listed set of tools. A support-scoped agent can list and update tickets without ever seeing admin tooling. Connect AI Gateway enforces this through <strong>Workspaces</strong>, which <a href="https://www.cdata.com/blog/mcp-field-level-security">scope access to specific data assets</a>, so a compromised agent’s reach stays limited to what that workspace exposes, plus role-based restrictions layered on top of the source system’s own controls.</p><h3>RBAC inheritance from source systems</h3><p>Role-based access restricts tool access by role, so marketing sees marketing tools and finance sees financial systems. The stronger model enforces the permissions your CRM, ERP, or database already has, at execution time, so governed access carries over to AI without a second permission model to maintain.</p><p>Connect AI’s passthrough identity works this way: permission changes in the source system apply to AI access automatically, since the gateway never holds the authoritative record.</p><h3>SIEM-exportable audit logs and governance</h3><p>Every request passing through the gateway gets logged, producing one trail instead of fragments scattered across every connected system. Connect AI Gateway logs each request down to the user, agent, timestamp, and data volume, exportable in SIEM-compatible formats for whatever platform your team already runs. A security sign-off typically checks four things: SSO integration, runtime identity propagation, least-privilege scoping, and <a href="https://www.cdata.com/blog/ai-model-audit-trail-guide">logs your SIEM can ingest</a>.</p><h3>Centralized governance vs. ad hoc per-user MCP configurations</h3><p>Centralized governance means agents connect once to a single governed endpoint that routes to the right backends behind it. The alternative is scattered, per-user access configurations that no one can see in one place and no one fully controls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6sZCXDfl6YyW_OC2ZkCBIA.png" /></figure><p>Identity providers like Microsoft Entra push lifecycle events through SCIM, so a new hire added to a group picks up the right tool access on the provider’s normal provisioning cycle, typically landing within about 40 minutes rather than requiring someone to configure each server by hand.</p><h3>Deployment patterns for private, on-premises, and hybrid environments</h3><p>Some teams need managed SaaS, some need a private cloud connector running inside their own network, and some need both depending on the system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uvj6ZIElpBxXHtJNoIOo-A.png" /></figure><p>Connect AI’s <a href="https://docs.cloud.cdata.com/en/Connect-Gateway">Private Cloud Connector</a> runs as a Docker container or Kubernetes deployment inside your infrastructure and opens an outbound-only connection back to Connect AI. On-premises systems are never directly reachable from the public internet, and nobody needs to open inbound firewall rules to use them.</p><h3>Enable governed AI access without security debt with CData Connect AI</h3><p><a href="https://www.cdata.com/ai/">CData Connect AI</a> gives your agents governed, auditable access to the systems you already run, with passthrough identity, Workspaces, and SIEM-compatible logging built in.</p><p><a href="https://www.cdata.com/ai/signup/">Start a free trial of CData Connect AI</a></p><h3>Frequently asked questions</h3><p>What is an enterprise MCP gateway?</p><p>A control plane that sits between AI agents and MCP servers, handling authentication, routing, policy enforcement, and audit logging from one governed entry point instead of leaving each server to manage its own.</p><p>How is an enterprise MCP gateway different from an MCP server?</p><p>An MCP server connects to one system and exposes its tools. The gateway governs access across many MCP servers at once, adding authentication, role-based access, and a unified audit trail that no individual server provides on its own.</p><p>Is an enterprise MCP gateway the same thing as an API gateway?</p><p>No. An API gateway authenticates callers and routes traffic at the service level. An MCP gateway works at the tool level, propagating the identity of the person behind each agent action and enforcing per-client consent on individual tool calls, which is finer-grained than what an API gateway checks.</p><p>What security controls does an enterprise MCP gateway provide?</p><p>Authentication, role-based access control, per-client consent, token audience validation, rate limiting, IP restrictions, and audit logging exportable in SIEM-compatible formats.</p><p>Does running MCP behind the firewall require a gateway, or can we just use individual MCP servers?</p><p>Individual MCP servers have no built-in access control or audit trail. Without a gateway, every server needs its own auth setup and produces its own fragmented logs, which is the shadow IT pattern a gateway is built to prevent.</p><p>How does an enterprise MCP gateway support compliance and audit requirements?</p><p>Every request passing through the gateway is logged in one place, with the user, agent, and timestamp attached to each entry. That gives compliance teams one exportable record instead of reconstructing activity from scattered agent logs.</p><p>Can an enterprise MCP gateway connect AI agents like Copilot Studio or Claude to on-premises data sources?</p><p>Yes. CData Connect AI documents direct MCP integrations with both Microsoft Copilot Studio and Claude. For on-premises sources, the CData private cloud connector runs inside your network and opens an outbound-only connection, so agents reach internal systems without inbound firewall rules.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/enterprise-mcp-gateway-behind-firewall"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2f75d508477" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/enterprise-mcp-gateway-securely-run-mcp-behind-the-firewall-e2f75d508477">Enterprise MCP Gateway: Securely Run MCP Behind the Firewall</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP for Enterprise Knowledge Management: Real-Time Control]]></title>
            <link>https://medium.com/cdata-software/mcp-for-enterprise-knowledge-management-real-time-control-74d0c562bff2?source=rss----47213ef81ae5---4</link>
            <guid isPermaLink="false">https://medium.com/p/74d0c562bff2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-integration]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[CData Software]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 13:53:45 GMT</pubDate>
            <atom:updated>2026-11-05T13:58:17.935Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/300/0*DyEEj4KhTPtScO1X" /></figure><p>An AI assistant can only answer from the knowledge it can reach. In most companies, that knowledge sits in support ticket histories, HR policy libraries, product documentation, contract repositories, and data warehouses, each with its own permissions. The <a href="https://www.cdata.com/blog/what-is-mcp">Model Context Protocol (MCP)</a> gives AI a standard way to query those systems live instead of relying on copies. The challenge for IT is control: who can query what, under whose identity, with what record of it. As an MCP platform, <a href="https://www.cdata.com/ai/">CData Connect AI</a> answers that question by providing AI agents governed, live access to hundreds of enterprise data sources.</p><h3>What is MCP for enterprise knowledge management?</h3><p>The MCP is an open standard that defines one uniform way for AI applications to connect to external data and tools. <a href="https://www.anthropic.com/news/model-context-protocol">Released by Anthropic in late 2024</a> and now governed by the <strong>Linux Foundation’s Agentic AI Foundation</strong>, it gives AI assistants and agents live, permission-aware access to the systems that hold enterprise knowledge. For <strong>knowledge management</strong> <strong>(KM)</strong>, this changes the model: instead of copying content into a central repository for AI to search, the AI queries each system of record the moment someone asks a question. Knowledge stays where it’s governed, and answers reflect the current record, not the last sync.</p><h3>Why traditional knowledge management systems fall short for AI</h3><p>Enterprise KM usually fails at access control and freshness, not data collection. KM platforms, content management tools, and centralized warehouses are good at gathering content, but what they hand to AI is a snapshot, synced on a schedule, that the model then treats as current truth. Gartner predicts that through 2026, <a href="https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk">organizations will abandon 60% of AI projects unsupported by AI-ready data</a>. The failures trace to data readiness and governance, not to the models. Traditional data management runs on reporting cycles, such as quarterly audits and monthly content reviews, while production AI consumes data continuously.</p><p>Consider an employee asking the internal assistant about the travel policy, if HR revised the document after the last sync, the model answers from outdated content. The model did its job. The failure happened upstream.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q5-399lZBitaJal86Z2C0Q.png" /></figure><h3>How MCP architecture works: servers, clients, and hosts</h3><p><a href="https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture">MCP’s architecture</a> defines three roles: <strong>host, client, </strong>and<strong> server</strong> as summarized in the table below. What matters for governance is the 1:1 mapping between client and server as each client holds a dedicated, isolated connection to one server, so state from one server never leaks into another, and every connection can be permissioned and observed on its own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V1jJkZcm5jSfgCNlthHMRQ.png" /></figure><p>MCP servers run in cloud, on-premises, or hybrid environments, matching enterprises where documentation, ticketing, and legacy systems sit in different places. Building a separate server per system multiplies the work, which is why a <a href="https://www.cdata.com/ai/">managed platform</a> that presents many governed sources through one standardized interface has become the common enterprise pattern.</p><h3>Real-time data access vs. static, batch-synced knowledge repositories</h3><p>MCP queries the live system of record the moment a question comes in. A batch-synced repository answers from a snapshot that ages until the next sync. That difference decides answer quality like when an assistant reads current enterprise data instead of static content, its responses reflect what’s true currently, which directly reduces wrong answers. For teams deploying copilots and agents, this is the fix for the upstream failure described earlier as the problem was never the model, but the outdated data grounding it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AEnkv8UZc1KgxckS1rqE4A.png" /></figure><p>MCP doesn’t replace <strong>retrieval-augmented generation (RAG)</strong>. <a href="https://www.cdata.com/blog/rags-to-riches-enterprise-ai-context-strategies-part1">RAG determines what context reaches the model</a>, while MCP standardizes how that context gets accessed, which is why retrieval and agent workflows often query enterprise systems through MCP.</p><h3>The governance layer: workspace isolation, passthrough permissions, and audit trails</h3><p>Governance is usually framed as a security add-on, but for MCP-based KM it’s the feature. The same interface should enforce least-privilege access for every caller, whether AI agent, BI tool, or human analyst, reusing the identity infrastructure the enterprise already runs. The protocol supports this directly: the <a href="https://modelcontextprotocol.io/extensions/auth/enterprise-managed-authorization">MCP Enterprise-Managed Authorization extension</a> puts the enterprise <strong>identity provider (IdP)</strong> between MCP clients and servers, so the organization controls access through its existing <strong>single sign-on (SSO)</strong> with <strong>OpenID Connect</strong> or <strong>Security Assertion Markup Language (SAML)</strong>. This enables passthrough permissions: an agent’s queries inherit the end user’s identity, so the agent sees only what that user is authorized to see.</p><p>These controls are prerequisites because MCP without strong identity, role-based access control, and observability becomes unmanaged integrations, and the failure modes are documented: over-privileged agents, shared credentials multiplying across service accounts, and audit blind spots where no one can say which user ran a query. Connect AI counters each one through a single governed control plane. Passthrough OAuth/SAML identity runs every query under the end user’s permissions, workspace isolation limits each agent to its authorized systems, and query-level audit logs export to your <strong>security information and event management (SIEM)</strong> system. IT extends the SSO infrastructure it already operates to AI assistants, with no new access path and no new audit gap.</p><h3>Which enterprise environments benefit most from MCP-based knowledge management</h3><p>MCP-based KM fits best where three conditions already hold: governed IT with an SSO identity provider, knowledge spread across multiple systems, and AI already deployed. Passthrough extends the existing identity to agents with no new access path, MCP exposes each system as a governed source whether it runs in cloud or on-premises, and live grounding improves the tools, such as <strong>Copilot</strong>, <strong>Claude</strong>, or custom agents, that employees already query. Compliance obligations strengthen the case further, because query-level attribution produces the records auditors ask for.</p><p>Gartner advises classifying agents by autonomy level rather than treating trust as binary, and MCP’s scoped, per-server access supports that: each agent’s reach can match its autonomy. Some environments gain less. A team with one data source or no identity infrastructure can serve AI well with simpler retrieval over static content. The strongest fit is mid-to-large enterprises with governed IT and many sources, the profile a managed control plane serves by unifying hundreds of sources under one identity-first interface.</p><p>The <a href="https://www.cdata.com/blog/enterprise-mcp-use-cases-cio-guide">CIO guide to enterprise MCP use cases</a> maps the fit by organizational context.</p><h3>Operationalizing real-time control with a managed MCP platform</h3><p>When IT doesn’t provide a governed path to enterprise knowledge, employees route around it with unapproved AI tools. A managed MCP platform gives IT a way to approve AI access while enforcing identity, least privilege, and auditability on every query. Adoption works best in stages. Pilot one high-value knowledge source, such as internal documentation, and validate against real user questions that:</p><ul><li>Passthrough identity resolves correctly for each user</li><li>Permissions deny what they should deny</li><li>Retrieval quality holds on live data</li></ul><p>Then expand source by source, since each new system inherits the same identity and audit configuration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UrWfnJ_yP0Ep5SVJKIc2Og.png" /></figure><h3>Frequently asked questions</h3><p>How is MCP different from traditional enterprise KM platforms and content management systems?</p><p>KM platforms and content management systems store and organize content. MCP is a connectivity protocol that gives AI live, governed access to those systems, so answers reflect current records instead of the last sync.</p><p>Does MCP work with the enterprise knowledge bases and information systems we already have?</p><p>Yes. MCP servers expose existing platforms such as SharePoint, ServiceNow, and SAP, and CData Connect AI provides prebuilt, governed connectivity to hundreds of sources through one MCP endpoint.</p><p>How does MCP for enterprise KM prevent AI hallucinations?</p><p>It grounds answers in live, authoritative enterprise data instead of static snapshots, so responses reflect what is true at request time rather than what was true at the last sync.</p><p>How does MCP handle knowledge silos across multiple enterprise systems?</p><p>Each system becomes a governed MCP source that AI queries through one standard interface. Data and permissions stay in place, and the assistant answers from live records across systems.</p><p>What’s the difference between MCP and RAG for enterprise KM?</p><p>RAG retrieves relevant content into a model’s context. MCP is a live access layer that complements RAG by exposing the systems that retrieval and agent workflows query.</p><h3>Build governed, real-time AI access with CData Connect AI</h3><p>MCP works for enterprise KM only when control is real-time. That means passthrough identity, workspace isolation, and an audit trail on every query. With <a href="https://www.cdata.com/ai/">CData Connect AI</a>, your AI tools query hundreds of enterprise systems live through single governed platform.</p><p><em>Originally published at </em><a href="https://www.cdata.com/blog/mcp-enterprise-knowledge-management"><em>https://www.cdata.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74d0c562bff2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/cdata-software/mcp-for-enterprise-knowledge-management-real-time-control-74d0c562bff2">MCP for Enterprise Knowledge Management: Real-Time Control</a> was originally published in <a href="https://medium.com/cdata-software">CData Software</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>