Claude Code Router: Route Claude Code to Any Model (2026 Guide)

What Claude Code Router Does

Claude Code talks Anthropic's message format and, by default, only to Claude. Claude Code Router (@musistudio/claude-code-router) is an open-source proxy that intercepts those requests, rewrites them for whatever provider you point it at, and picks the model per request type. You launch the agent with ccr code instead of claude, and nothing else about your workflow changes.

The reason to bother is cost shape. A coding session is not one workload. Summarizing a diff, naming a branch, and compacting old context are throwaway tasks. Planning a refactor is a reasoning task. Applying a 40-line edit across three files is where quality actually matters. Paying frontier-model output rates for all of it is the default, and it is wasteful.

Install in 60 Seconds

Two global npm packages: Claude Code itself, and the router.

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

The CLI is ccr. The commands you will actually use:

The config.json Structure

Configuration lives in ~/.claude-code-router/config.json. Two blocks matter: Providers (an array of where models come from) and Router (which model handles which request type). A minimal working file:

{
  "Providers": [
    {
      "name": "deepseek",
      "api_base_url": "https://api.deepseek.com/chat/completions",
      "api_key": "sk-xxx",
      "models": ["deepseek-chat", "deepseek-reasoner"],
      "transformer": { "use": ["deepseek"] }
    }
  ],
  "Router": {
    "default": "deepseek,deepseek-chat",
    "background": "deepseek,deepseek-chat",
    "think": "deepseek,deepseek-reasoner",
    "longContext": "deepseek,deepseek-chat",
    "longContextThreshold": 60000,
    "webSearch": "deepseek,deepseek-chat"
  }
}

Each provider entry needs four fields: name (a unique label you reference later), api_base_url (the full endpoint, not just the host), api_key, and models (the model IDs that provider serves). transformer is optional and covered below. Optional top-level fields include APIKEY to lock down the local router, PROXY_URL, API_TIMEOUT_MS, and HOST.

The Five Routing Rules (Plus Two)

The Router object is where the savings live. Each key takes a provider,model string. Claude Code Router inspects every outbound request and picks the model by these rules:

The single highest-leverage rule is background. Claude Code fires background requests constantly to summarize and compact, and they do not need a frontier model. Pointing background at a small or open-source model removes a large chunk of token spend that was invisible before.

Adding Providers

A provider is one API endpoint and the models behind it. Add as many as you want; the Router rules mix and match across them. DeepSeek and OpenRouter as a two-provider setup:

{
  "Providers": [
    {
      "name": "openrouter",
      "api_base_url": "https://openrouter.ai/api/v1/chat/completions",
      "api_key": "sk-or-xxx",
      "models": ["anthropic/claude-sonnet-4.6", "google/gemini-3.1-pro"],
      "transformer": { "use": ["openrouter"] }
    },
    {
      "name": "deepseek",
      "api_base_url": "https://api.deepseek.com/chat/completions",
      "api_key": "sk-xxx",
      "models": ["deepseek-chat", "deepseek-reasoner"],
      "transformer": { "use": ["deepseek"] }
    }
  ],
  "Router": {
    "default": "openrouter,anthropic/claude-sonnet-4.6",
    "background": "deepseek,deepseek-chat",
    "think": "deepseek,deepseek-reasoner",
    "longContext": "openrouter,google/gemini-3.1-pro",
    "webSearch": "openrouter,anthropic/claude-sonnet-4.6"
  }
}

For local models, point a provider at Ollama's OpenAI-compatible endpoint (http://localhost:11434/v1/chat/completions) and list the model tags you have pulled. See running Claude Code with Ollama for the local-model caveats around tool calling.

Routing to Open-Source Models

Any OpenAI-compatible endpoint serving open weights drops in as a single provider. Morph hosts Qwen 3.5, Qwen 3.6, MiniMax M2.7, and DeepSeek V4 Flash on one endpoint, so they all live under a single provider entry:

{
  "name": "morph",
  "api_base_url": "https://api.morphllm.com/v1/chat/completions",
  "api_key": "sk-...",
  "models": [
    "morph-qwen35-397b",
    "morph-qwen36-27b",
    "morph-minimax27-230b",
    "morph-dsv4flash"
  ]
}

That maps cleanly onto the router rules: a small model on background, a large MoE on default, and a long-context model where the conversation grows. Why open weights at all is its own decision; the cost and benchmark case is in the best open-source coding model and per-model API pages.

Dynamic Switching and Subagent Routing

The static Router rules are defaults, not handcuffs. Two overrides:

Per-subagent routing pairs naturally with how multi-agent systems already split work. Cheap parallel readers feeding a single expensive writer is the same shape as agent orchestration generally.

Transformers: Making Any API Speak Claude

Not every provider accepts Anthropic-shaped requests. Transformers are the adapters that reshape requests and responses so a given API behaves. The built-ins cover the common cases:

Apply them per provider with "transformer": { "use": ["deepseek"] }, or per model inside a provider. You can also load a custom transformer from a JavaScript module for an endpoint nothing built-in covers.

The One Thing That Breaks: Edit Accuracy

Routing changes which model writes the code. It does not change Claude Code's edit format, which was tuned for Claude's exact output style. Route the default model to DeepSeek, Qwen, or GLM and you will see more failed edits: wrong line numbers, dropped context lines, malformed search-and-replace hunks. The agent retries, burns tokens, and sometimes corrupts the file.

The root cause is asking one model to both decide the change and emit a byte-perfect diff. Splitting those jobs fixes it. A fast-apply model takes the rough edit plus the original file and merges them deterministically, so the routed model only has to describe the change, not format it.

Frequently Asked Questions