Claude Code Auto Mode on AWS Bedrock, Vertex, and Foundry — Enable It, Route It, Audit It

On May 30, Anthropic shipped a single change in Claude Code that quietly reshapes how enterprises route coding agent workloads: auto-mode now runs on AWS Bedrock, Google Vertex AI, and Microsoft Foundry for Opus 4.7 and Opus 4.8. Until now, auto-mode — Claude Code's background, permission-free operation mode — was exclusive to the Anthropic API. The expansion turns every team's "which provider?" question from a theoretical debate into an operational decision with real billing, latency, and compliance consequences.

What happened

On May 30, 2026, Claude Code v2.1.15x introduced a single configuration toggle: set CLAUDE_CODE_ENABLE_AUTO_MODE=1 and auto-mode works through Bedrock, Vertex, and Foundry for both Opus 4.7 and Opus 4.8. The feature had been Anthropic-API-only since its introduction in Week 20 (May 11–15).

The same release window (May 29–30) also brought plugin auto-loading from .claude/skills directories and claude plugin init scaffolding — but the auto-mode expansion across cloud platforms is the infrastructure story that changes routing architecture.

Auto-mode itself is Claude Code's mechanism for running coding agents in the background without permission prompts: the agent checks out branches, edits files, runs tests, and opens PRs while you're away. Before May 30, teams that wanted auto-mode had to use the Anthropic API directly, even if their organization's primary cloud commitment was with AWS, Google, or Microsoft.

Why it matters for AI engineering teams

This isn't a cosmetic "more platforms" checkbox. It changes three things that directly affect engineering budgets and architecture:

1. Billing aggregation. If your organization already has an AWS Enterprise Discount Program, Google Committed Use Discount, or Microsoft Azure Consumption Commitment, routing Claude Code through Bedrock/Vertex/Foundry means Opus 4.8 consumption counts toward your existing cloud spend — not a separate Anthropic API invoice. For teams burning hundreds of dollars per day on coding agents, this can be the difference between Opus fitting within existing procurement or requiring a net-new vendor approval cycle.

2. Data residency and VPC compliance. Bedrock, Vertex, and Foundry each operate within your cloud account's network boundary. Auto-mode workloads that touch proprietary source code no longer need to cross the public internet to reach api.anthropic.com. For teams in regulated industries — finance, healthcare, defense — this removes a blocker that made auto-mode non-viable even when the feature itself was desirable.

3. Latency and availability surface. Each provider path has different characteristics. Bedrock in us-east-1 ≠ Vertex in us-central1 ≠ Foundry in East US ≠ Anthropic API. A team that runs Claude Code sessions from CI runners in AWS will see different round-trip times on Bedrock versus Anthropic API. And each platform has its own availability SLA — routing auto-mode through your primary cloud provider means one less external dependency in the critical path.

The router/operator angle

For teams that operate an AI gateway or routing layer — including TheRouter users — this change introduces a new routing dimension: provider path is now a first-class variable alongside model selection.

Here's the decision matrix that didn't exist before May 30:

The operational implications cascade:

• Fallback chains now have a fourth path. If Anthropic API has an outage or rate-limit spike, an operator can fall back to Bedrock or Vertex for the same model. This isn't model fallback — it's provider-platform fallback, and it works even when the model itself (Opus 4.8) isn't changing.
• Cost routing gets more granular. Teams can now route based on which cloud commitment pool is underutilized this month. If AWS spend is below the EDP threshold, route to Bedrock. If Google CUD is maxed, route to Anthropic API.
• Observability splits across providers. Each platform exposes different logging, metrics, and audit trails. Bedrock logs go to CloudTrail/CloudWatch; Vertex to Cloud Logging; Foundry to Azure Monitor. A routing layer that doesn't normalize these into a single observability view loses the ability to compare performance across paths.

What to watch: The CLAUDE_CODE_ENABLE_AUTO_MODE=1 flag is a single environment variable. There is no per-provider toggle — it's all-or-nothing across Bedrock, Vertex, and Foundry. Anthropic may introduce per-platform toggles if enterprise customers request different policies for different cloud providers.

What TheRouter users should watch or try

1. Test latency from your CI/CD infrastructure. Spin up identical Claude Code auto-mode sessions through each available provider path and measure end-to-end time from task submission to PR creation. The results will surprise you — cloud inter-region latency often dominates model inference time.

2. Build a provider-path fallback policy. If your team relies on auto-mode for overnight coding agent runs, a single-provider dependency is now unnecessary. Configure a routing rule that tries Bedrock first (if you're an AWS shop), falls back to Anthropic API, and alerts if both fail.

3. Audit your billing aggregation. Calculate whether routing Claude Code through an existing cloud commitment changes your effective per-token cost. Bedrock and Vertex pricing for Opus 4.8 may differ from direct Anthropic API pricing, and the delta multiplied by hundreds of thousands of daily tokens adds up fast.

The auto-mode expansion is a signal: Anthropic is treating cloud platform distribution as a first-class deployment model, not an afterthought. Teams that build their routing architecture to exploit this — rather than treating the Anthropic API as the only path to Claude Code — will have lower latency, simpler procurement, and one fewer single point of failure.