Claude Code Model Router

🔀 ccmr — Claude Code Model Router

Switch between 40+ third-party models mid-conversation — no restarts, no config editing.

ccmr is a lightweight gateway that runs on your machine: Claude Code talks to it, and it forwards every request verbatim to whichever model you pick — DeepSeek, Kimi, Qwen, GLM, MiniMax, Step, MiMo or Doubao Seed, 44 models across 8 vendors, all prewired. Type /model kimi in a session and the very next message runs on Kimi. When a key dies, the error tells you exactly which env var to fix, where to get a new key, and the command that verifies it. Config edits hot-reload in about a second. Loopback-only by default; everything stays local.

English · 简体中文

Why ccmr

For most setups, switching models means editing env vars or config files and restarting Claude Code. ccmr is a live gateway: switching happens inside the session (/model, effective on the next message, context preserved), because Claude Code always faces one stable local address — the change happens behind it.

Honest scope: ccmr is an Anthropic-protocol passthrough gateway — it does no protocol conversion and only talks to vendors' official Anthropic-compatible endpoints (that's why Zhipu CN is supported via its Coding Plan only: its pay-as-you-go side offers no Anthropic endpoint). Streaming, tool calls and thinking blocks are forwarded as-is; your context is never parsed or rewritten.

Quick start
You can also run the gateway in the foreground with ccmr start (logs in your terminal); gateways auto-started by ccmr claude log to ~/.ccmr/gateway.log.
Plain claude still uses your official subscription — the two modes have fully isolated configs.

Commands
Config discovery: -c path > ./models.yaml > ./config/models.yaml > ./.claude-router.yaml > ~/.ccmr/models.yaml. An explicit -c that…