Source: ai-research/claude-blog-claude-apps-gateway.md — Anthropic first-party blog post, claude.com/blog/introducing-the-claude-apps-gateway (fetched 2026-07-16). Refreshed 2026-09-29 with gateway entries from the Claude Code v2.1.251–v2.1.284 release notes (see frontmatter).
Anthropic shipped the Claude apps gateway: a self-hosted control plane for running Claude Code on Amazon Bedrock and Google Cloud, giving enterprises corporate SSO login, centrally enforced managed settings, role-based access, and per-user cost attribution without hand-rolling that infrastructure themselves. It directly answers the operational cost of the “bring your own cloud” deployment model — previously an org running Claude Code on Bedrock or GCP had to provision a cloud credential per developer, manually push settings to every laptop, and stand up separate tooling just to see per-developer spend. The gateway folds all three of those jobs into one container Anthropic builds and ships.
Key Takeaways
- It’s a single stateless container, not a new agent or SDK. Deployed on Linux, backed by a PostgreSQL database. It holds the org’s upstream cloud credential, authenticates developers against the org’s own identity provider (IdP), distributes and enforces managed settings, and reports per-user usage to a collector the org operates.
- Onboarding/offboarding collapses to IdP membership. Add a developer to the IdP to onboard them; remove them to offboard. No per-developer credential provisioning step to remember.
- Built into the same
claudebinary, not a bolt-on proxy. Because the gateway and the CLI client are built together by Anthropic, the/loginflow is gateway-aware, the client applies managed settings automatically at sign-in, and policy is enforced consistently on every request — there’s no separate wrapper script or shim to maintain. - Inference traffic stays off Anthropic’s servers by default. The gateway does not send inference traffic or usage data to Anthropic unless the org explicitly configures it to use the Claude API instead of Bedrock/GCP — the whole point is that inference already routes through the org’s existing cloud relationship.
- Anthropic is publishing the protocol. Stated explicitly so other gateway developers can implement the same features — a signal this is meant to be an open integration point, not a closed proprietary tunnel.
- Configuration surface named in the post:
gateway.yamlfor gateway config,forceLoginMethod/forceLoginGatewayUrlsettings so the client always authenticates through the gateway, andmanaged-settings.jsonfor the centrally-enforced policy the gateway distributes. - Available now, with a getting-started doc at
code.claude.com/docs/en/claude-apps-gateway.
What Problem This Solves
The post frames the “before” state as three separate manual burdens for Bedrock/GCP customers running Claude Code at scale:
- Credential provisioning — a cloud credential per developer, set up by hand.
- Settings distribution — manually pushing configuration to every laptop, with no central enforcement.
- Spend visibility — standing up separate tooling just to see what each developer is costing.
The gateway replaces all three with one deployment: IdP-based auth (no per-developer credential step), gateway-enforced managed-settings.json (no manual laptop-by-laptop pushes), and a usage collector the org operates (spend visibility without bespoke tooling). This is the Bedrock/GCP-specific counterpart to what Anthropic’s admin-analytics post (published the same week) ships for orgs on Anthropic’s own hosted infrastructure — same underlying goals (policy enforcement, per-user cost attribution), different deployment target.
How It Fits the Existing Governance Model
This is infrastructure-level enforcement, not prompt-level instruction — the same category of control Agent Guardrails argues for generally: a PreToolUse hook or managed settings, not a CLAUDE.md instruction, is the right tool for an absolute must, because prompted instructions can fail under pressure or via injection. The gateway takes that same logic to the identity/access layer: SSO and centrally enforced managed-settings.json are policy an individual developer’s local settings can’t override, distributed and enforced by infrastructure the org — not the developer’s laptop — controls. It’s also a direct instantiation of the never trust, always verify and least privilege principles in Anthropic’s Zero Trust for AI Agents eBook: every request authenticates against the org’s own IdP regardless of where it originates, and role-based access scopes what each developer can do rather than granting broad standing access.
Try It
- If your org already runs Claude Code on Bedrock or Google Cloud, read the getting-started doc (
code.claude.com/docs/en/claude-apps-gateway) before your next credential-provisioning cycle — the gateway may eliminate the per-developer credential step entirely. - Check whether your
managed-settings.jsonpolicy is currently enforced centrally or just distributed by convention. The gateway’s value is specifically that policy is enforced consistently on every request, not just pushed once and trusted to stick. - If you’re evaluating build-vs-buy for a self-hosted spend-attribution tool for Claude Code, note that Anthropic is publishing the gateway’s protocol — worth checking whether an existing internal tool can integrate against it instead of being replaced by it.
Release-note additions (Claude Code v2.1.251–v2.1.284, August–September 2026)
The gateway ships inside the claude binary, so its changes arrive in Claude Code release notes. Digests: W35 · W36 · W37 · W38 · W39 · W40.
Cost and spend.
pricing:ingateway.yamlflows to signed-in clients through managed settings, so/costand telemetry match the spend meter (v2.1.268).modelPricingand the gatewaypricingblock accept amultiplierabove 1, up to 10, for internal chargeback markups (v2.1.271).- A Spend limit bar in
/usageand arate_limits.spend_limitstatus-line field (v2.1.251); from v2.1.284 the limit shows in dollars (e.g. “500.00 spent this month”) andrate_limits.spend_limitgainsused_usd,limit_usdandperiod.
Sign-in and identity.
forceLoginGatewayUrlmakes machines gateway sessions from startup, likeforceLoginMethod: "gateway"(v2.1.265); machines pinned toforceLoginMethod: "gateway"ignore a leftover API key or claude.ai login (v2.1.261).oidc.scope_on_refreshfor identity providers that return an id_token on refresh only when asked foropenidagain (v2.1.260); certificate client authentication (private_key_jwt) with the identity provider (v2.1.284).gatewayInternalNetworkslets admins allow/loginto a gateway on the organization’s own public IPv4 block (v2.1.268).- Request headers up to 256 KiB are accepted, so sign-ins whose IdP lists many groups no longer get 431 on every request (v2.1.284).
Upstreams and operations.
- Bedrock upstreams gain
assume_role(call Bedrock as an IAM role assumed through STS, optionally one session per developer) andguardrail: {id, version}to apply an Amazon Bedrock guardrail to every request (v2.1.281); amantleupstream provider for Bedrock’s Mantle endpoint (v2.1.283). CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1for gateways whose only egress is a forward proxy, and an upstreamheaders:map for static headers to a proxy you run (v2.1.277).CLAUDE_GATEWAY_DRAIN_TIMEOUT_MS(in-flight requests finish for up to 25 seconds on SIGTERM) andstore.connect_timeout_seconds(v2.1.274);store.readiness_grace_secondskeeps/readyzready through a short Postgres outage (v2.1.282);load_test_modebuilds and signs requests without sending them upstream (v2.1.283).telemetry.resource_attributes(v2.1.281);auth: { google: {} }exports telemetry straight to Google Cloud’s OTLP endpoint (v2.1.284).- Client hint headers:
CLAUDE_CODE_GATEWAY_HINT_HEADERS=1opts in tox-claude-code-request-class,-agent-type,-prev-tool-durations,-compactionand-context-compacted(v2.1.273), plusx-claude-code-prompt-idto group the requests serving one user prompt (v2.1.283).
Models and policy.
- In gateway sessions
fableandbestkept resolving to Fable 5 “for now” after Fable 5.1 launched, because gateways not yet configured for it reject it (v2.1.257). - Startup warnings when a managed policy’s
availableModelsis empty or leaves out the model Claude Code starts on (v2.1.284); the gateway marks each 1M-capable model so Claude Desktop offers the 1M option (v2.1.284). desktoppolicy blocks accept newer Claude Desktop keys such asuserPluginMarketplacesEnabledanduserPluginUploadsEnabled(v2.1.260).- Auto mode defaults to the server-side classifier on gateways (v2.1.278) — see Auto Mode.
Release-note additions (Claude Code v2.1.316, 2026-10-31)
- Pooled group budgets:
pooled_budgeton the spend-limits admin API sets one monthly budget that a group’s members share, on top of each member’s own limit. - Layered policies:
managed.policy_mode: layeredapplies every policy that matches a developer, each merged over the ones listed below it, instead of only the first match. - Bedrock guardrail input tagging:
guardrail.input_taggingon Bedrock upstreams wraps prompts in the input tags a Bedrock guardrail’s prompt-attacks filter needs;all_messages_and_systemalso tags the system prompt. - Model checks fixed: the gateway now checks every model a request names against
availableModels,deniedModelsandmodels. A request naming a blocked model gets a 400 that names the field. - Sign-in:
claude auth loginno longer refuses to run where managed settings require the gateway; in a terminal it shows the gateway sign-in. The VS Code chat panel can sign in to the gateway without a terminal. - Full release: What’s New — Week 44.
Open Questions
- The post’s own “How the gateway works” section states “The gateway handles:” followed by what was evidently a structured bullet list in the live page — the list did not survive extraction as text (a common failure mode for accordion/icon-list components on claude.com). The “Deploying the gateway” paragraph is the reliable paraphrase; the exact original bullet wording is not reproduced here to avoid inventing it.
- No pricing, licensing, or self-hosting resource-requirement detail (CPU/memory sizing for the PostgreSQL-backed container) is given in the post.
- The page’s own “Related posts” carousel names four companion announcements — all four now resolved: “Claude Cowork is coming to mobile and web” (folded into
claude-cowork.md/cowork-getting-started.md), “Bringing Claude Code and Claude Cowork to government” (new article), “Auto mode for Claude Code” (new hub article, 2026-07-16), and “Making Claude Cowork ready for enterprise” (folded intoclaude-cowork.md’s existing Apr 8 timeline row + a new Customer case studies section, 2026-07-16). - Whether the gateway supports platforms beyond Bedrock and Google Cloud (e.g., Azure) is not addressed — the post scopes explicitly to those two.
Related
- Agent Guardrails: Hooks, Permissions, and Sandboxing Patterns — the general framework this gateway instantiates at the identity/access layer: deterministic, infrastructure-enforced policy over prompted instruction
- Zero Trust for AI Agents (Anthropic eBook) — the never-trust/always-verify and least-privilege principles this gateway’s SSO + role-based access directly implements
- Managed Agents — Self-Hosted Sandboxes + MCP Tunnels — the parallel self-hosted-infrastructure pattern for Managed Agents (customer-controlled execution environment) that this gateway extends to Claude Code’s identity/policy layer
- Giving Admins More Visibility and Control Over Claude Spend — the contemporaneous hosted-side counterpart; per-user cost attribution here pairs with the admin spend-analytics dashboard there
- Claude Code CLI Reference — the gateway is built into the same
claudebinary this reference documents;/login,gateway.yaml, andmanaged-settings.jsonare CLI-level surfaces - Managed Agents — the broader Anthropic-hosted-agents primitive this gateway’s self-hosted-control-plane model contrasts with