Claude Code Router: the trick to cheaper Claude Code (with the trade-off nobody mentions)

TL;DR npm install -g @musistudio/claude-code-router puts a local proxy between Claude Code and the model, letting you route requests to DeepSeek, Qwen, Ollama, OpenRouter, or whatever. Configure ~/.claude-code-router/config.json with your providers and rules (per scenario: default, background, think, longContext, webSearch), launch with ccr code, and switch models on the fly with /model. The honest trade-off nobody mentions: you keep Claude Code's UX, not Claude. DeepSeek inside Claude Code is still DeepSeek — better than ChatGPT at many tasks, but worse than Sonnet 4.6 at agentic work and reasoning.

More and more searches lead to "claude code free", "claude code ollama", "claude code unlimited". Behind all of them is the same question: can I keep coding with this CLI when my plan runs out, or when the code shouldn't leave my machine? The community's answer — not Anthropic's — is Claude Code Router, an open-source project with 34k+ stars that intercepts Claude Code's calls and translates them to almost any provider.

It's not magic. It's a proxy that changes ANTHROPIC_BASE_URL to point at your machine (defaults to http://127.0.0.1:3456) and rewrites requests to the format of whichever provider you pick. What you keep: the CLI, slash commands, hooks, skills, plugins. What you don't keep: Claude itself.

Result:

> ccr code

Claude Code Router running on localhost:3456
Default model: deepseek,deepseek-chat

> /model openrouter,anthropic/claude-sonnet-4-6

Switched to claude-sonnet-4-6 via OpenRouter for this turn

> /model ollama,qwen2.5-coder:14b

Switched to qwen2.5-coder:14b (local) for this turn

Install

# 1. Claude Code (if you don't have it)
npm install -g @anthropic-ai/claude-code

# 2. The router
npm install -g @musistudio/claude-code-router

# 3. Verify
ccr --version

Minimal config (DeepSeek + Ollama)

Create ~/.claude-code-router/config.json:

{
  "Providers": [
    {
      "name": "deepseek",
      "api_base_url": "https://api.deepseek.com/chat/completions",
      "api_key": "sk-...",
      "models": ["deepseek-chat", "deepseek-reasoner"]
    },
    {
      "name": "ollama",
      "api_base_url": "http://localhost:11434/v1/chat/completions",
      "api_key": "ollama",
      "models": ["qwen2.5-coder:14b", "llama3.1:8b"]
    }
  ],
  "Router": {
    "default": "deepseek,deepseek-chat",
    "background": "ollama,qwen2.5-coder:14b",
    "think": "deepseek,deepseek-reasoner",
    "longContext": "deepseek,deepseek-chat",
    "webSearch": "deepseek,deepseek-chat"
  }
}

⚠️ Heads-up on the DeepSeek models (correct as of publication): deepseek-chat and deepseek-reasoner are scheduled for retirement on July 24, 2026, per the DeepSeek API official docs. The official replacements are deepseek-v4-flash and deepseek-v4-pro. If you land on this tip after that date, swap the names in models and Router — the router syntax doesn't change, only the model identifier.

The five Router keys aren't decoration — they're distinct scenarios Claude Code triggers based on the kind of operation:

Launch with ccr code (instead of claude). Hot-reload config: ccr restart. There's a UI: ccr ui. To swap models mid-conversation: /model <provider>,<model> (no space after the comma).

The honest trade-off

What you won't find in "Claude Code unlimited" blogs:

• You keep the UX, not Claude. DeepSeek-Chat inside Claude Code is still DeepSeek — good for refactors and boilerplate, worse at agentic work, complex tool use, and multi-step reasoning.
• Some features need Claude. Auto mode needs Anthropic's classifier. /advisor calls Opus specifically. The Figma + Chrome MCP combo with /goal works because Claude handles the agentic loop — swap in a weaker model and it breaks within 3-4 turns.
• It's a community project. When Claude Code's API changes, or when a provider deprecates models (like DeepSeek with v4), the router can break until musistudio patches. This isn't production-critical infrastructure.
• Privacy depends on the provider. Ollama → your machine, 100% local. DeepSeek/Qwen → the provider's servers (Chinese in both cases). Decide based on the repo.

When it IS worth it

• You ran out of Pro/Max plan and need to keep going today → switch to DeepSeek for a few hours
• Sensitive code that can't leave your machine → route to local Ollama
• Massive cheap tasks (rename 200 files, generate boilerplate) → background to a cheap model
• You live in a region without direct Anthropic billing → OpenRouter via the router
• You want to compare models inside the same flow → swap on the fly with /model

When NOT to use it

• When you're doing complex agentic work — auto mode, agent teams, subagents. Those flows assume Claude-specific capabilities.
• When your bottleneck is reasoning, not cost. Paying for Sonnet 4.6 and planning well is cheaper than saving on tokens and rewriting badly.
• When you're saving tokens, not budget. If the problem is your context filling up, the answer is 10 habits to save tokens, not switching models.

Reference

Official docs: musistudio/claude-code-router (GitHub) · Routing scenarios docs · DeepSeek API Docs

Requirements

• Node.js 18+
• Claude Code v2.0+
• An API key from your chosen provider (DeepSeek, OpenRouter…) or Ollama running locally