claude-code-router

Claude Code Router (ccr) is a request-layer router that sits between Claude Code and one or more model providers. The Claude Code binary is unchanged; ccr intercepts API calls, applies a routing table, transforms requests/responses per provider, and forwards. Anthropic's UX, anyone else's model.

What it does

• Multi-provider — OpenRouter, DeepSeek, Gemini, Ollama, Volcengine, SiliconFlow, ModelScope, DashScope, AIHubMix, plus arbitrary OpenAI-compatible endpoints.
• Routing rules — default, background, think, longContext, image, webSearch slots each map to a provider,model pair, so different request types go to different models.
• Transformers — per-provider request/response munging (the openrouter, deepseek, gemini, enhancetool, maxtoken, reasoning, tooluse transformers ship in-box; new ones can be plugged in).
• Dynamic switching — /model slash command from inside Claude Code rewrites the default route on the fly, and ccr model does the same from a separate terminal.
• CLI — ccr code, ccr model, ccr status, ccr logs, ccr restart.
• GitHub Actions integration — non-interactive mode (NON_INTERACTIVE_MODE: true) sets CI=true, FORCE_COLOR=0, fixes stdin handling so Claude Code can run in a workflow.
• Env-var interpolation — "$OPENAI_API_KEY" / "${GEMINI_API_KEY}" in config.json so secrets stay out of the file.
• Plugin system — drop a TypeScript file into the transformer directory, ccr loads it on restart.

How it works

config.json (at ~/.claude-code-router/config.json) declares providers and a router. Each provider has name, api_base_url, api_key, models, and optional transformer. The Router section picks defaults:

{
  "Providers": [
    {
      "name": "deepseek",
      "api_base_url": "https://api.deepseek.com/chat/completions",
      "api_key": "$DEEPSEEK_API_KEY",
      "models": ["deepseek-chat", "deepseek-reasoner"],
      "transformer": {
        "use": ["deepseek"],
        "deepseek-chat": { "use": ["tooluse"] }
      }
    },
    {
      "name": "ollama",
      "api_base_url": "http://localhost:11434/v1/chat/completions",
      "api_key": "ollama",
      "models": ["qwen2.5-coder:latest"]
    }
  ],
  "Router": {
    "default": "deepseek,deepseek-chat",
    "background": "ollama,qwen2.5-coder:latest",
    "think": "deepseek,deepseek-reasoner"
  }
}

Two log surfaces: pino-formatted server logs at ~/.claude-code-router/logs/ccr-*.log (HTTP / API events) and a separate claude-code-router.log for routing decisions and business-logic events. APIKEY adds a shared-secret check in the Authorization / x-api-key header; if unset, HOST is forced to 127.0.0.1 so an unauthenticated daemon can never bind to the network.

Install

npm install -g @anthropic-ai/claude-code
npm install -g @musistudio/claude-code-router

Then point Claude Code at ccr (the project's docs walk through the env vars).

Where it fits

CCR is the request-layer counterpart to cc-mirror — same goal (Claude Code on non-Anthropic models), different mechanism. CC-Mirror clones the install per provider; ccr keeps one install and routes at request time. Both ship a ccrouter integration so cc-mirror variants can use ccr as their backend. cursor-bridge is the opposite end of the same family: no configuration, no daemon, one backend, and a process that lives and dies with the terminal.

Conceptually it's an llm-api-routing-layer — same shape as Vercel AI Gateway, Helicone, LiteLLM — except the upstream client is Claude Code specifically, and the transformers know about Claude Code's tool-call format. The 33k-star count makes it the most-adopted Claude-Code-shaped LLM gateway.

Limitations

• Configuration is text-edited JSON — the surface is big and easy to mis-shape.
• Tool-use fidelity across providers is uneven; transformers paper over the worst gaps but model-specific quirks remain.
• Routing costs the prompt cache. KV state is model-specific and does not travel, so a /model switch or a slot that sends think requests elsewhere re-prefills the whole session at uncached prices — see prompt-caching-in-agents. portal-shunt-token-routing routes by task instead: bulk reads go out as separate calls, so the main session's cache stays intact.
• A single ccr instance is single-tenant by default — the APIKEY mechanism is a minimum, not full multi-tenant auth.
• Sponsored by Z.ai (the GLM-CODING-PLAN promo banner sits at the top of the README); their interest in keeping ccr useful is structural, not contractual.

Repo: https://github.com/musistudio/claude-code-router (MIT, 33.8k stars).