Amazon Bedrock Claude Code: setup, model access, and limits

What changes when Claude Code runs through Bedrock

Claude Code remains a coding agent. Bedrock supplies the model endpoint and AWS authentication path. Anthropic's Bedrock guide documents the required Bedrock switch and AWS region; AWS credentials or IAM roles authorize model access. Claude Code still runs commands on its host; using Bedrock does not put every tool, local file, or connected service inside AWS.

This can fit teams that already govern model usage through AWS. It also means the AWS account owner must grant the model access and the runtime permission deliberately. Do not paste long-lived keys into a shell profile or repository. Prefer the credential mechanism your organization already audits.

Set up and verify the model path

First, submit the Anthropic-model use-case form in the Bedrock Model catalog once for the AWS account, then confirm the intended model is enabled in the intended region. For an individual, run claude, choose 3rd-party platform > Amazon Bedrock, and use the wizard. It can use an AWS profile, Bedrock API key, access-key credentials, or environment credentials, verify accessible models, and save the configuration. For CI or a managed rollout, Claude Code uses the default AWS SDK credential chain: configure an AWS profile or other approved credential source, then set CLAUDE_CODE_USE_BEDROCK=1. Set AWS_REGION only to override the active profile’s region or when no profile region exists; run /status to see the resolved region.

Grant IAM permissions for the approved model and inference profile, rather than copy broad or long-lived credentials into a repository. For a team rollout, pin explicit Bedrock model or inference-profile IDs after checking availability. Anthropic’s guide shows ANTHROPIC_MODEL='us.anthropic.claude-sonnet-4-6' as an inference-profile example, but valid IDs depend on the AWS account and region. A base model ID from an old guide can fail with a routing error.

Validate with a small repository task before using production code. Ask Claude Code to explain one non-sensitive module and run an existing focused test. Check /status for the resolved region and provider, then check the selected model, any AWS permission error, and the test result. Treat a successful explanation as connectivity evidence, not proof that the agent is safe to modify a production service. Bedrock WebSearch is unavailable on this path, so verify required feature availability before committing to it.

A manual SSO example

For an approved AWS SSO profile, Anthropic documents this route. Replace your-profile-name with a profile your administrator configured:

aws sso login --profile=your-profile-name
export AWS_PROFILE=your-profile-name
export CLAUDE_CODE_USE_BEDROCK=1
claude

In the session, check /status and confirm the selected model. Current region resolution checks AWS_REGION, AWS_DEFAULT_REGION, the active profile, then us-east-1. Explicitly approve the resolved region rather than relying on the fallback. The wizard can also be reopened with /setup-bedrock.

For a fleet, review the documented IAM configuration with your AWS administrator. Invocation needs bedrock:InvokeModel and streaming access through bedrock:InvokeModelWithResponseStream; inference-profile discovery and resolution use ListInferenceProfiles and GetInferenceProfile. Limit resources to the intended models/profiles where supported. Do not copy a broad sample policy unchanged into production.

Diagnose the failing layer

Changing a permission prompt in Claude Code cannot fix an AWS IAM denial. Likewise, a successful model call does not prove that repository changes or later eesel actions are authorized.

Pricing and limits to check

Use the current AWS Bedrock pricing page for the selected model, region, and inference mode. Do not reuse a Claude 3-era price table for a current model or assume that a batch discount applies to an interactive Claude Code session. Model input and output usage are only part of the cost. Budget for retries, long codebase context, and engineering review. Check AWS quotas and observed throttling on the workload you expect, not only a single prompt.

AWS and Anthropic also document different feature availability across Bedrock and Anthropic-operated AWS offerings. If you need a particular API feature, beta capability, or agent skill, verify it on the Bedrock route before committing to the architecture.

Use Bedrock-backed Claude Code to review an eesel teammate

Once Claude Code is configured, you can ask it to use eesel CLI for a support task alongside your engineering work. For example: “Compare our approved release note with the support teammate's instructions and propose a correction to its troubleshooting guidance.”

The coding agent calls the CLI and reads its JSON output. A person can run the same commands directly, or put approved inspection commands in a script. They operate the same eesel teammate and workspace as the dashboard—not a copy hosted in Bedrock.

After signing in to the intended eesel workspace, inspect an existing teammate:

npx @eesel/cli whoami
npx @eesel/cli status --agent "Product support"
npx @eesel/cli instructions --agent "Product support"

Node.js 18.17+ is required. Confirm the agent name and owner-approved access before sharing the returned content with a coding agent. AWS authentication does not authorize eesel, and the data the coding agent reads may enter its model context. Review that data flow under your organization's policy.

Suppose an approved release note says a renamed account setting is now under Preferences, while the teammate's instructions still refer to Profile. Ask Claude Code to identify the stale sentence and propose the smallest correction, citing the supplied release note. It should flag missing evidence rather than guess a UI location.

The support owner reviews the proposed diff. Check instructions --help for the edit syntax and preview a supported write with --dry-run before applying it. Inspect the saved instructions through the CLI or dashboard. Then, after checking action permissions, test a synthetic question about the renamed setting and a case involving an older product version. Chat is real, billed work; inspect activity and test actual helpdesk delivery separately.

This keeps the responsibilities visible: AWS governs Claude Code's model access, eesel governs its workspace teammate, and connected systems retain their own permissions. Try eesel