A developer in a Berlin fintech stares at the screen, prompts flying into Claude Code, but extended thinking? Dead on arrival because Bedrock’s lurking underneath.
Claude Code on Amazon Bedrock. That’s the phrase echoing through enterprise chats right now — teams chasing data control without ditching Anthropic’s coding wizard. But here’s the rub: direct Bedrock hookup cripples the tool. No extended thinking for those multi-file refactors. No web search for fresh bugs. It’s like buying a sports car and yanking the turbo.
And yet, companies flock there. Why?
Why Bedrock Beats Anthropic Direct for Locked-Down Teams
Data sovereignty seals it. Your prompts, responses, logs — they never escape your AWS perimeter. Encryption? Yours to tune. Retention? You decide. Regulated sectors — finance, healthcare — they can’t touch Anthropic’s API without compliance nightmares.
Pick your region. eu-west-1 for Europe? Done. Route through specific accounts. Bedrock self-services that, no begging Anthropic support.
Your IAM policies, VPCs, CloudTrail? Bedrock slots right in. No new security circus.
Pricing’s pure on-demand tokens. Anthropic? Slap a $20/user/month base on top, one-year lock-in. Scaling teams hate commitments.
But Claude Code detects Bedrock — sets CLAUDE_CODE_USE_BEDROCK=1 — and flips switches off. Client-side sabotage.
With Bedrock, all your data, logs, and metrics live inside your AWS account. You control retention, encryption, and access. Your prompts and responses don’t leave your infrastructure boundary.
That’s the siren song. Straight from the source.
What Gets Gutted — And Why It Hurts
Extended thinking. Gone. That’s Claude’s secret sauce for wrestling thorny codebases — reasoning step-by-step across files, spotting subtle bugs. Without it? You’re on vanilla inference, good but not great.
Web search. Poof. Training cutoff means recent libs, new errors? Claude’s blind. One fresh Stack Overflow hit changes everything.
Tool use? Partial at best.
Here’s the scorecard:
| Feature | Direct Bedrock | Anthropic API |
|---|---|---|
| Extended thinking | Disabled (client-side) | Enabled |
| Web search | Disabled | Enabled |
| Tool use | Partial | Full |
Bedrock supports these natively. Claude Code just won’t send the requests. Dumb.
Teams grumble. Productivity dips 20-30% on complex tasks, I’ve heard from devs switching.
The Gateway Hack: Pretend It’s Anthropic
Workaround? Slap a proxy in front. Speaks Anthropic Messages API. Claude Code thinks it’s home free — fires all features. Proxy translates to Bedrock’s dialect.
LiteLLM works, sorta. AWS has a reference arch, but it’s deploy-heavy.
Then there’s Claude Code AWS Gateway (CCAG). Open-source, Rust binary — 15MB slim. MIT licensed. Purpose-built.
It maps models region-smart. Sanitizes cache_control. Filters beta headers. Normalizes SSE streams. Generic proxies choke; this doesn’t.
Postgres backend for state. Self-host it.
Suddenly: Bedrock perks intact. Full Claude Code firepower.
But CCAG goes further — team-scale goodies neither vendor matches.
Budgets per-user, per-team. Throttle at limits. Webhooks, SNS alerts.
OIDC SSO. Okta, Entra, Google — log in smoothly. Ditch AWS creds.
Multi-account routing. Pool quotas across regions, dodge 429s. Fallbacks. Isolate by country.
Admin dashboard. Spend graphs. Team mgmt. One-command dev setup.
SCIM 2.0. Auto-sync users.
Is CCAG the New Standard for Enterprise AI Coding?
Look, this echoes early OpenAI days. Devs built proxies to stitch models, dodge rate limits. NGINX for LLMs. CCAG’s that — but laser-focused.
My bet? Forks explode. Enterprises fork private, add bells. Anthropic notices, bundles similar (too late). Bedrock teams standardize here. It’s the architectural shift: proxies aren’t hacks; they’re the glue for sovereign AI.
Corporate spin calls Bedrock ‘enterprise-ready.’ Sure, for basics. But full Claude Code? They sandbagged. CCAG exposes the gap — and fills it gratis.
Teams I’ve pinged love it. One at a bank: “Extended thinking back online, spend capped, SSO instant.” Boom.
Short history lesson: Remember Kubernetes? Started as proxy-ish orchestrator for containers. Now infra king. Gateways like this? Same trajectory for AI tools.
Why Does This Matter for Your Dev Team?
If you’re greenfield, Anthropic direct shines — full features, co-work bundled.
Enterprise? Bedrock + gateway. Sovereignty without handcuffs.
Scaling? Those budget controls save black holes. Multi-region? Latency wins, quota pools kill flakes.
Skeptical? Deploy the binary. Test extended thinking on your repo. See the diff.
It’s not hype. It’s the how: Rust efficiency, Postgres reliability, AWS-native. Why: Because AI coding demands unclipped wings, even in cages.
One caveat — you’re self-managing. Updates manual. But MIT open? Community fills gaps fast.
🧬 Related Insights
- Read more: Pyroscope and Alloy Slice Through TON Blockchain Bottlenecks
- Read more: 6 Backend Truths AI Skips: Why Your Spring Boot App Crashes Tuesdays at 2 AM
Frequently Asked Questions
What is Claude Code AWS Gateway?
CCAG is a lightweight Rust proxy that lets Claude Code use all features on Bedrock by mimicking Anthropic’s API.
Does Bedrock support extended thinking for Claude Code?
Bedrock does, but direct Claude Code disables it client-side — gateways fix that.
Is CCAG free and open source?
Yes, MIT licensed, single binary, no tiers.
