Start with the supported connection
Atlassian's client guide explicitly supports ChatGPT. OpenAI also documents its Atlassian Rovo connector, introduced in December 2025 for eligible accounts. Current availability still depends on your account and workspace controls. The requirements include an Atlassian Cloud site, access to the chosen client, and an authorized authentication method. Do not treat a self-managed Confluence Server installation as the same setup.
Rovo MCP connects compatible AI clients to Atlassian information and tools. It can search and summarize Confluence content and expose supported page or work-item changes. Available tools and permissions determine what a particular connection can do.
A practical setup sequence is:
- Ask the workspace owner whether Atlassian access is enabled for your role.
- Find the Atlassian Rovo app or plugin on the supported ChatGPT surface. Follow its current connection prompts rather than an old screenshot's menu labels.
- Authenticate to the intended Atlassian site and inspect the requested access.
- Start with a page you are authorized to read and a question with a known answer.
- Open the returned source link and compare the actual page before relying on the summary.
OpenAI's plugin guide explains that installation can make bundled capabilities available to new chats, while a connector may still need separate authentication. Installation, provider access, and permission to perform actions are different steps.
For a first check, use a bounded request: “Find the approved onboarding checklist in the People space. List its required steps and link to the source. Do not create or edit anything.” The wording states your intent; configure actual read-only access where needed.
Compare the routes without confusing the products
| Route | Useful for | What you must manage |
|---|---|---|
| Atlassian Rovo connection in ChatGPT | Search or summarize authorized wiki content during a chat | Client availability, authentication, tool permissions, source checks |
| Exported page or document | A small, approved snapshot for one analysis | Upload eligibility, copy permissions, age of the content |
| A custom integration | An application with requirements the available connector does not meet | API contract, authentication, retrieval, authorization, maintenance |
| Rovo inside Confluence | Questions and collaboration without leaving the wiki | Atlassian access, source quality, usage terms |
| eesel teammate with Confluence knowledge | Repeated support questions in an existing support channel | Selected knowledge, teammate instructions, escalation and outcome review |
Atlassian documents Rovo search inside Confluence as an alternative to switching to ChatGPT. It answers from accessible content and links to sources. That is Rovo, not a ChatGPT connection hidden behind a different interface.
A custom GPT action and a custom application are also different projects. Do not automatically add OpenAI API charges to every GPT action: determine which service actually makes the model calls and how that product is billed.
Treat permissions and freshness as separate checks
An authorized user may still be blocked by organization policy. Atlassian's Rovo MCP access controls can restrict access beyond ordinary page permissions. That documentation also distinguishes OAuth policy enforcement from API-token authentication; do not assume the same control covers both.
OpenAI's workspace app controls distinguish app availability, allowed actions, provider authorization, and runtime permissions. Review those layers before enabling a write-capable tool. A read-only trial does not justify later enabling all actions.
Freshness is a different question. An exported PDF does not update when its source page changes. A connection that retrieves content on demand is not automatically an indexed copy of every page. Where sync is offered, inspect its actual scope and status rather than promising “always up to date.”
Use two test users with different approved access if you are evaluating an organization rollout. Have the owner verify that each sees the intended material and cannot retrieve a restricted test page. Do not use a real confidential document merely to see whether access controls fail.
Use Confluence knowledge in an eesel support teammate
The eesel Confluence integration reads selected page text with app View access. Its self-serve Marketplace setup is for Confluence Cloud; Data Center needs a separate managed discussion. It cannot write to the wiki, and page attachments are not indexed.
Choose pages deliberately, then check that indexing has completed. Updates are picked up automatically, with roughly hourly reconciliation, not an unconditional real-time guarantee. A stale answer needs investigation before an instruction change.
That separation matters for support. If a teammate cannot find the escalation procedure, the problem might be a missing source rather than poor wording. Fixing instructions will not make an unindexed attachment searchable.
Inspect an escalation answer with eesel CLI
Suppose your Confluence playbook says a confirmed P1 outage goes to the Incident Lead, while an unconfirmed report goes to normal support investigation. A support owner wants to check whether the teammate distinguishes those cases.
The official CLI works from a human's terminal, a script, or coding agents such as Claude Code, Codex, and Cursor. It returns JSON from the same eesel teammate and workspace used in the dashboard. Node.js 18.17 or newer is required.
After signing in to the intended workspace, replace support with its actual teammate:
npx @eesel/cli integrations --agent support
npx @eesel/cli integrations download list --agent support
npx @eesel/cli instructions --agent support
Ask the coding agent to compare that evidence with the approved playbook: is the right source present, is it ready, and does a standing rule contradict it? Confirm selected pages and app access in Confluence as well. The CLI connection list is not a complete permission audit.
Before chatting, the owner checks all enabled actions in the dashboard and connected helpdesk. Restrict unwanted writes or outbound contact. Confluence's read-only integration does not make every other connected service read-only.
Then use this complete fictional case:
npx @eesel/cli new --name "confluence-escalation-check" --agent support
npx @eesel/cli chat "Fictional evaluation. Approved playbook: confirmed P1 outages go to the Incident Lead; unconfirmed reports go to support investigation. Customer reports one failed login. Status: no outage confirmed. Draft the next reply and proposed handoff, identify any conflict with standing instructions, and do not declare a P1 incident. Do not contact anyone, change configuration or records, or approve actions." --agent support
npx @eesel/cli activity --agent support
Chat is real billed work; fictional facts and a no-action prompt are not a sandbox. Review whether the reply acknowledges the report without declaring an outage or inventing evidence.
If instructions are the problem, have the coding agent propose an exact correction for owner approval. If the source is missing or stale, repair that source setup first. For a supported configuration write, consult command help and preview the actual proposed write with --dry-run.
After approval, retest in fresh conversations using the standing policy rather than supplying it in the prompt. Include a confirmed outage and an unconfirmed report, then verify the native helpdesk handoff with safe test tickets. This connects knowledge access to a support outcome you can actually inspect.
Budget for the chosen route
Separate the ChatGPT plan, Atlassian usage, any custom application's API and hosting costs, and eesel workspace work. Atlassian's Rovo usage guidance distinguishes free single-product context lookups from enriched context calls, including unified search, that consume credits. Other AI intelligence events can also consume credits. Check the exact tools your workflow uses rather than assuming all MCP calls are free or have the same cost.
For eesel, inspect the workspace's current billing state and budget for real chat tests. Avoid using an old subscription table as a quote for a different product or plan.
Choose where the answer belongs
For a person's wiki research, begin with an authorized Atlassian connection in ChatGPT or Rovo inside Confluence. For a support job, connect approved knowledge to a teammate and verify its sources, rules, and handoff.
Try eesel to review that support workflow through the dashboard or CLI. Start with one playbook and two cases whose expected outcomes you already know.
Frequently asked questions
Can ChatGPT read private Confluence pages?
Yes, through an authorized connection such as Atlassian Rovo MCP when your account and workspace permit it. Merely pasting a private URL does not give ChatGPT your Confluence session or permissions.
Do I need a custom GPT to connect Confluence?
Not necessarily. Atlassian supports ChatGPT as a client for Rovo MCP. Check the available Atlassian app or plugin before deciding to build a custom API integration.
Does an exported PDF stay synchronized?
No. It is a copy from the export time. Later page edits and permission changes do not automatically update that copy, so manage its replacement and sharing separately.
Does every Confluence connection allow writes?
No. Rovo MCP can expose write tools subject to permissions and controls. The eesel Confluence integration is a read-only knowledge source, with no Confluence actions or triggers.
Why might an eesel teammate miss a Confluence answer?
Check selected pages, app View permissions, indexing status, and whether the information exists only in an attachment. The integration indexes selected page text, not attachments.
