Claude Code Router: run Claude Code on other models

What the tool actually is

Claude Code Router runs a small server on your machine, by default on http://127.0.0.1:3456, that speaks the Anthropic Messages API on the front and whatever the target provider speaks on the back. Claude Code is pointed at it and behaves normally; every turn is translated, routed, and forwarded. A management UI runs separately on http://127.0.0.1:3458.

# CLI, requires Node.js 22+
npm install -g @musistudio/claude-code-router
ccr ui        # opens the management interface

# or the desktop app from GitHub releases
# or Docker
docker compose up -d --build

The project has grown well beyond its name. Its own description in August 2026 is "one local control plane for every AI agent", and it now fronts Codex, Grok CLI, and Kimi CLI as well as Claude Code, with credential pools and profile switching. That is worth knowing before you evaluate it: you are adopting a control plane, not a shim.

How its routing works

Configuration lives at ~/.claude-code-router/config.json and has two blocks that matter: Providers, an array describing where models come from, and Router, which maps request classes onto them.

{
  "Router": {
    "default": "openrouter,anthropic/claude-sonnet-5",
    "background": "openrouter,google/gemini-flash",
    "think": "openrouter,anthropic/claude-opus-5",
    "longContext": "openrouter,google/gemini-pro",
    "longContextThreshold": 60000
  }
}

The honest risk list

• It is unofficial, and it sits in the middle of your credentials. A local proxy holding provider keys is a supply-chain surface. Read the release you install, pin the version, and do not run it on a machine where a compromise would be expensive.
• Two moving API shapes. The tool translates between what Claude Code emits and what each provider accepts. Both change. 1,090 open issues is not an indictment of the project, it is an accurate picture of how much surface that translation has.
• Non-Anthropic models in Claude Code are not the same product. Claude Code's prompting, tool definitions, and agent loop are tuned for Claude. A model that benchmarks well elsewhere can behave noticeably worse inside this harness, especially on multi-step tool use, and the failure looks like the tool being broken rather than the model being a poor fit. Check a candidate against our coding-model leaderboard before you slot it into default, and treat the score as a floor rather than a promise inside this harness.
• Terms of service deserve five minutes of your own reading. Anthropic documents base-URL overrides and third-party endpoints such as Bedrock and Vertex, so a custom endpoint is a supported configuration. What is a different question is routing an Anthropic subscription credential through third-party software, or using a subscription rail for automated traffic it was not sold for. Check your own agreement rather than a forum post.
• No support contract. When it breaks mid-sprint, you are reading TypeScript. For a personal machine that is fine. For twenty engineers it is a bus factor.

The four alternatives, in order of how little you have to run

1. Just the environment variables

Claude Code reads its endpoint from the environment. If your target already serves an Anthropic-shaped API, there is nothing to install and no process to keep alive.

export ANTHROPIC_BASE_URL="https://your-gateway.example.com"
export ANTHROPIC_AUTH_TOKEN="your-key"
export ANTHROPIC_MODEL="claude-opus-5"
export ANTHROPIC_SMALL_FAST_MODEL="claude-sonnet-5"

claude

Inside the session, /status prints the base URL in use, which is the fastest way to prove the variables reached the process. The catch is that this gives you one model, not four routing classes.

2. LiteLLM in front

LiteLLM serves an Anthropic-compatible surface as well as an OpenAI one, so it can be the thing ANTHROPIC_BASE_URL points at while translating to any of a hundred providers behind it. You get real fallbacks, budgets, and logs, and you get a service to operate. See the self-hosting guide.

3. OpenRouter

One key, hundreds of models, nothing to run, automatic provider fallback. You pay 5.5% on credit purchases (minimum $0.80, checked August 2026) and your prompts cross a third party. Fine for personal work, a procurement conversation for a company.

4. A maintained hosted endpoint

Questions people ask

Keep reading

Continue through LLM gateways and routing with LLM failover: surviving a provider outage without losing the turn.