Claude Code talks to Amazon Bedrock with one setting, and Codex has been on Bedrock since June 2026. How to set either up for a whole team (SSO, pinned models, prompt caching, guardrails, per-team costs) and the mistakes that make the bill or the security review go badly.
A coding agent reads your source, runs commands and sends large chunks of code to a model. Through a vendor subscription, that traffic goes to the vendor's API under accounts you manage separately. Through Bedrock, it goes to a model endpoint in your AWS account: IAM decides who may call which models, CloudTrail records every call, Guardrails can filter content, and the spend lands on the bill you already reconcile. For many companies that turns a three-month security review into a configuration change.
In a dedicated AWS account — Anthropic's documentation recommends one, for cost tracking and access control — submit the Anthropic use-case form once. With AWS Organizations, the management account can do it for every child account through the PutUseCaseForModelAccess API.
Give engineers short-lived credentials through IAM Identity Center: aws sso login, then AWS_PROFILE. Claude Code's awsAuthRefresh setting can run the login for them when credentials expire.
Set CLAUDE_CODE_USE_BEDROCK=1 and AWS_REGION. Engineers can also use the /setup-bedrock wizard, which writes the settings for them.
Grant a narrow IAM policy: bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:ListInferenceProfiles and bedrock:GetInferenceProfile on your inference profiles and foundation models.
Pin model versions in managed settings — ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL — so the whole team moves to a new model when you decide, not when a default changes.
Book a free 30-minute call. You'll leave knowing what we'd do first and roughly what it costs, whether or not you hire us.
Application inference profiles per team, mapped with modelOverrides, so the cost dashboard shows who spends what. Budgets with alerts on each.
ENABLE_PROMPT_CACHING_1H for long sessions on large repositories — the 1-hour cache costs more to write than the 5-minute default, and saves far more when the same context is read all afternoon.
Guardrails, attached with the X-Amzn-Bedrock-GuardrailIdentifier and GuardrailVersion headers through ANTHROPIC_CUSTOM_HEADERS, where your policy requires content filtering.
ANTHROPIC_BEDROCK_SERVICE_TIER, where the model supports it — flex for unattended batch work, priority for interactive use under load.
availableModels with enforceAvailableModels in managed settings, so nobody routes around your choices.
AWS made Codex generally available on Bedrock on June 1, 2026 — the Codex app, CLI and IDE integrations, with all inference routed through Bedrock and billed per token, with no seat licences or per-developer commitments. The same account, Identity Center sign-in and cost tracking work for both, so teams can use whichever agent suits their work without a second procurement.
Which repositories agents may work in, and whether production credentials may ever be present in their environment (the answer should be no).
Which commands run without approval — tests, linters, builds — and which need a human: anything that deploys, deletes or touches infrastructure.
That agent-written code goes through the same review as anyone else's, and who is accountable for it (the engineer who merged it).
Once a month: what changed in tech that month and what it means for you, from cloud and AI to security and the tools you build with. Every item links to its source. No sales pitch, unsubscribe in one click.
Book a free 30-minute call about it. Bring your setup and your bill; you'll leave knowing what we'd do first and roughly what it costs, whether or not you hire us.
Or keep reading
Bedrock vs OpenAI for Enterprise in 2026: Which Models, Which Contract, Where the Data Stays
OpenAI's models have run on Amazon Bedrock since June 2026, so the platform decision is no longer about which models you can use. What it is about now (data handling, identity, lifecycle and lock-in) and a short plan for teams that built on OpenAI directly.
Read
Amazon Bedrock AgentCore in 2026: What Each Piece Does and Which Ones You Need
AgentCore went generally available in October 2025 and kept growing: Runtime, Gateway, Identity, Memory, Observability, Policy, harness, payments. What each does, which problems it solves that your framework does not, and a sensible order to adopt them.
Read
As of October 25, 2026. Provided as is, for general information, and accurate to the best of our knowledge on the date shown. Not legal, compliance, tax, financial or other professional advice. Prices, limits and vendor services change; check current vendor documentation before relying on anything here. Terms of use.