Claude Apps Gateway
Anthropic quietly ends
Claude's wiring problem for the enterprise.
Anthropic's new Claude apps gateway folds SSO, spend caps, and revocation for Claude Code on Amazon Bedrock and Google Cloud into a single self-hosted container. Deprovision a user in your IdP and their access expires within the hour. The daily grind of handing out credentials — and clawing them back — quietly disappears.
The Bottleneck
Why shipping Claude Code
used to eat half a day
Deploying Claude Code on Amazon Bedrock or Google Cloud always hit the same wall. Every developer needed a Bedrock or Vertex IAM credential minted, baked into their laptop, and — every time they moved teams or left the company — manually revoked. Because those keys tie into the Anthropic session, disabling the user in your IdP wasn't enough: API calls kept flowing. Anthropic's June 29, 2026 announcement folds that daily chore into a single container.
The other pain was cost visibility. Nobody could tell which developer had run what against which model without back-solving from Bedrock or GCP invoices, or trawling CloudTrail. Spend caps weren't even possible. DevOps.com put it plainly: "IT doesn't fear handing Claude Code to developers — IT fears what happens after." The gateway attacks both frictions — key distribution and post-hoc billing — at once.
| Before (direct) | Via Claude Apps Gateway |
|---|---|
| Per-developer Bedrock / GCP IAM credentials | Corporate SSO login, 1-hour short-lived tokens |
| Manual revocation on offboarding, prone to gaps | Disable in your IdP → gone within session lifetime |
| Spend caps and model limits built ad-hoc | Caps at org / group / user, out of the box |
| Inference cost visible only after invoices | Per-user usage and spend, aggregated live |
Handing out keys is easy.
What enterprises really want is "pull one lever, they're gone."
How It Works
An OIDC relay
folded into one container
The gateway acts as an OIDC relying party for Google Workspace, Microsoft Entra ID, Okta or any standards-compliant IdP, and hands short-lived bearer tokens to Claude Code.
Sign in with corporate SSO
The client baked into the Claude Code binary hits your IdP over OIDC and receives a 1-hour bearer token from the gateway. Nobody has to store personal Bedrock or GCP credentials on their laptop anymore.
Gateway routes inference
The gateway forwards each request — by policy — to Claude on Bedrock, Google Cloud, or Microsoft Foundry, and fails over automatically when one path is down. Developers don't see the swap.
Disable in your IdP, gone in an hour
Off-boarding is a single IdP toggle. The next token refresh fails, and the current token expires within the session lifetime. No more ticketing Anthropic to revoke a departing developer.
Under the Hood
The control plane, folded
into one container
The gateway ships inside the same claude binary developers already install and runs as a single stateless container backed by Postgres.
Data path matters here. Anthropic's docs state plainly that unless the customer wires Claude API directly, "no inference traffic or usage data is sent back to Anthropic." A Bedrock customer's prompts never leave AWS, and even the gateway sends no telemetry home. That's the sentence compliance officers wanted to see.
Spend caps work at three tiers — org, group, and user — enforced pre-flight and alerted post-flight. InfoQ calls the gateway "the last missing piece for IT teams who want to let developers touch Bedrock Claude."
Who Benefits
Who this lands for
Same gateway, different wins by role.
IT / Security
SSO unification and instant revocation shrink the audit surface. Bedrock / GCP moves from "let them use it and hope" to "let them use it, and control it."
Finance / Leadership
Org-level spend caps kill the open-ended AI bill. Per-team usage is visible, so budget approvals get quicker.
Developers / Engineers
No more juggling personal tokens. Log in via SSO; if AWS or GCP goes wobbly, the gateway silently reroutes.
What's Next / Risk
Deployable, at last —
but there are traps left
With the gateway in place, "we can't safely hand out Claude Code" is no longer a defensible reason to postpone the rollout. The next question is what happens after distribution: prompt guidelines, audit-log retention, revocation drills. The gateway hands you a control plane; it doesn't hand you a policy.
There are three cautions. First, this deepens Anthropic lock-in. Unless OpenAI and Google ship an equivalent gateway for GPT and Gemini, multi-model comparison gets more friction, not less. Second, running Postgres in-house adds a real operational cost. If the gateway goes down, everyone loses login — that is a single point of failure that demands proper HA design. Third, spend caps are detect-and-alert, not pre-flight throttles: a runaway agent can still burn a lot of tokens inside a single cap window. Pair caps with rate limits, or you'll learn this the expensive way.
Three things to do next: (1) IT should run an IdP + revocation test as the PoC's first gate. (2) Finance should switch to a monthly per-user spend report cadence. (3) Engineering should chaos-test the gateway's failover behavior before any team depends on it. Do those three, and Bedrock Claude Code becomes a real SaaS to consume — not a security project to manage.