github.com/Clufy/claude-code-gateway · open source · MIT
Run Claude Code on any LLM with one Cloudflare Worker
claude-code-gateway speaks the Anthropic Messages API, so Claude Code talks to it as if it were Anthropic. Behind it, each model routes to open models on Cloudflare Workers AI, OpenAI, Gemini or Anthropic, and every session passes a security layer built for AI-assisted red and purple teaming.
Features
Multi-provider routing
Workers AI, OpenAI Chat Completions and Responses, Gemini and the Anthropic Messages API, with up to 3 fallback models per route.
Full Claude Code experience
Streaming, tool use, extended thinking, vision for flagged models, token counting and /v1/models with real context windows.
Secure by default
Required proxy key, provider host allowlist, body caps, rate limits and provider keys scrubbed from every response.
Session guard
Flags prompt injection, secrets and risky tool calls in each turn, with your own rules in GUARD_RULES.
Red-team harness
70 canary-based attack cases in 7 suites, mapped to the OWASP Top 10 for LLM Applications.
MCP server and purple loop
engagement, logs, redteam and purple tools at /mcp, plus Sigma rules drafted from findings and promoted to guard rules.
How it works
Claude Code ──Anthropic Messages API──▶ claude-code-gateway ──▶ Workers AI · OpenAI · Gemini · Anthropic
(your Worker)
auth · routing · fallbacks · session guard · scope check · /mcp
Quick start
Deploy with the button above, or from your machine:
npx wrangler login npm install && npm run check npm run deploy npx wrangler secret put PROXY_KEY
Then point Claude Code at the Worker in ~/.claude/settings.json:
{
"env": {
"ANTHROPIC_BASE_URL": "https://<worker-subdomain>.workers.dev",
"ANTHROPIC_CUSTOM_HEADERS": "x-api-key: your-proxy-key-value",
"ANTHROPIC_MODEL": "@cf/zai-org/glm-5.3",
"ANTHROPIC_SMALL_FAST_MODEL": "@cf/moonshotai/kimi-k2.6"
}
}
Full configuration, engagements, guard rules and the red-team harness are documented in the repository docs.
FAQ
Can I use Claude Code with other models?
Yes. Point ANTHROPIC_BASE_URL at your claude-code-gateway Worker and each model name routes to the backend you choose: Workers AI, OpenAI, Gemini, DeepSeek, Groq, OpenRouter or Anthropic.
Which open-source models does it ship with?
It ships configured for GLM-5.3, Kimi K2.6 and DeepSeek V4 Pro on Cloudflare Workers AI, with no GPU of your own.
Is the gateway secure by default?
A proxy key is required, provider hosts must be on an allowlist, bodies are size-capped, requests are rate limited and provider keys are scrubbed from every upstream response.
What does not work through the gateway?
WebSearch, prompt caching, the Files API, the Batches API and managed-agent features are not implemented.
How do I deploy it?
Use the one-click Deploy to Cloudflare button, or run npm install and npm run deploy with Wrangler, then set the PROXY_KEY secret.