Claude Code Router: Set Up OpenRouter, DeepSeek, and Ollama (2026 Guide)

The Claude Code Router is an open-source proxy tool that solves a specific limitation: Claude Code only supports Anthropic models natively. If you want to use Claude Code's interface with DeepSeek, GPT-4, Gemini, local Ollama models, or any of 200+ models available through OpenRouter, you need a router. The community claude-code-router project by @musistudio (28k+ GitHub stars, MIT license) does exactly this. It intercepts Claude Code's API requests and forwards them to whatever provider you configure.

This guide covers how to set up the community router step by step, which providers and models work best, how to troubleshoot common issues, and how Claude Code's built-in opusplan routing compares.

What Is Claude Code Router?

Claude Code Router is a local proxy server that sits between Claude Code and the AI model provider. Instead of sending requests directly to Anthropic's API, Claude Code sends them to the router running on your machine (default port 3456). The router transforms the request format and forwards it to whichever provider you've configured: OpenRouter, DeepSeek, Ollama, Gemini, or others.

How it works under the hood:

1. The router starts a local Express.js server on port 3456
2. It sets ANTHROPIC_BASE_URL to point at the local server
3. Claude Code sends API requests to the router instead of Anthropic
4. The router transforms the request format and forwards it to your configured provider
5. Responses stream back through the router to Claude Code

The result: you get Claude Code's interface, agentic capabilities, and workflow — but with whatever model you choose.

How to Set Up Claude Code Router

Step 1: Install

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

This installs the ccr command globally. You need Claude Code installed first (npm install -g @anthropic-ai/claude-code).

Step 2: Configure

Create ~/.claude-code-router/config.json with your providers and routing rules:

{
  "Providers": [
    {
      "name": "openrouter",
      "api_base_url": "https://openrouter.ai/api/v1/chat/completions",
      "api_key": "$OPENROUTER_API_KEY",
      "models": ["anthropic/claude-sonnet-4", "google/gemini-2.5-pro"],
      "transformer": { "use": ["openrouter"] }
    },
    {
      "name": "deepseek",
      "api_base_url": "https://api.deepseek.com/chat/completions",
      "api_key": "$DEEPSEEK_API_KEY",
      "models": ["deepseek-chat", "deepseek-reasoner"],
      "transformer": { "use": ["deepseek"] }
    },
    {
      "name": "ollama",
      "api_base_url": "http://localhost:11434/v1/chat/completions",
      "api_key": "ollama",
      "models": ["qwen2.5-coder:latest"]
    }
  ],
  "Router": {
    "default": "deepseek,deepseek-chat",
    "think": "openrouter,anthropic/claude-sonnet-4",
    "background": "ollama,qwen2.5-coder:latest"
  }
}

API keys can reference environment variables with $VAR_NAME syntax for secure key management.

Step 3: Run

The simplest way: start the router and launch Claude Code in one command:

ccr code

Or manage the service separately:

# Start the router in the background
ccr start

# Export environment variables for Claude Code
eval "$(ccr activate)"

# Run Claude Code normally — requests route through the proxy
claude

Supported Providers and Models

Recommendation for most developers: Start with OpenRouter as your single provider; it gives you access to every major model through one API key. Once you know which models you use most, add direct provider accounts (DeepSeek, Gemini) for lower per-token costs. Add Ollama if you want local/offline capability.

Context-Aware Routing Configuration

The Router config supports task-type routing. Each rule maps to a "provider,model" string, letting you automatically use the right model for each type of work:

This lets you use a cheap local model for background tasks, a reasoning-focused model for complex thinking, and a large-context model when your session grows, all automatically.

Troubleshooting Common Issues

Port 3456 already in use

Another process is using the router's default port. Either stop the conflicting process (lsof -i :3456 to find it) or configure a different port in your config.json with "port": 3457.

ANTHROPIC_BASE_URL override not working

If Claude Code still hits Anthropic's API after starting the router, check that the environment variable is set in the same shell session: run echo $ANTHROPIC_BASE_URL; it should show http://localhost:3456. Using ccr code (instead of manual ccr start + claude) handles this automatically.

Model doesn't support extended thinking

Not all models support Claude Code's extended thinking feature. If you see errors when the agent tries to "think," route think tasks to a model that supports it (Claude Sonnet/Opus via OpenRouter, or DeepSeek Reasoner) and use simpler models for default and background tasks only.

Tool use / function calling errors

Claude Code relies heavily on tool use (function calling). Some models have limited or no tool-use support. Stick to models known to handle tools well: Claude (any version), GPT-4+, Gemini Pro, DeepSeek Chat. Smaller local models via Ollama may struggle with complex tool-use patterns.

Ollama connection refused

Make sure Ollama is running (ollama serve) and the model is pulled (ollama pull qwen2.5-coder). The api_base_url must include the full path: http://localhost:11434/v1/chat/completions.

Built-In Alternative: opusplan

If you don't need non-Anthropic models, Claude Code has built-in model routing via the opusplan alias. It uses Opus 4.6 for planning and reasoning, then switches to Sonnet 4.6 for code execution, saving money on implementation tokens while keeping the best reasoning for architectural decisions.

Enable it with any of these methods (highest to lowest priority):

• Mid-session: /model opusplan
• At startup: claude --model opusplan
• Environment variable: export ANTHROPIC_MODEL=opusplan
• Settings file: Add "model": "opusplan" to ~/.claude/settings.json

Other available model aliases: sonnet, opus, haiku, sonnet[1m] (1M token context window). Aliases always point to the latest model version.

How to Configure Model Aliases

Beyond opusplan, you can pin what each alias resolves to using environment variables:

• ANTHROPIC_DEFAULT_OPUS_MODEL: what opus and opusplan (plan mode) resolve to
• ANTHROPIC_DEFAULT_SONNET_MODEL: what sonnet and opusplan (execution mode) resolve to
• ANTHROPIC_DEFAULT_HAIKU_MODEL: what haiku and background tasks resolve to

To pin a specific model version instead of using the alias, pass the full model ID: claude --model claude-opus-4-6.

Community Router vs opusplan: Which to Use?

Use opusplan if you want the simplest path to cost-optimized Claude Code usage with zero setup and full feature compatibility.

Use the community router if you want to use non-Anthropic models, run local models via Ollama, or need fine-grained control over which model handles which type of task.

Cost Tips for Claude Code

• opusplan for best quality-to-cost ratio: Opus-quality reasoning on architecture decisions, Sonnet efficiency on implementation. The recommended starting point for most developers.
• Haiku for quick tasks: use claude --model haiku for simple questions, quick edits, and boilerplate generation. Significantly cheaper than both Opus and Sonnet.
• Community router + cheap models for boilerplate: route background tasks to DeepSeek or local Ollama models. Save Anthropic credits for work that needs Claude's reasoning.
• Monitor your usage: use the /cost command in Claude Code to see token consumption and spending for your current session.
• Plan before executing: enter plan mode (Shift+Tab twice) to think through your approach before the model starts generating code. Better plans mean fewer wasted iterations.

For extending Claude Code's capabilities beyond model routing, browse available skills on PolySkill, portable packages that add coding standards, workflow automations, and specialized prompts to your sessions.

Related: How to Add Skills to Claude Code (2026) | Codex vs Claude Code: Full Comparison (2026) | Claude Code vs Cursor: Full Comparison (2026) | Claude Code vs Gemini CLI (2026)

Back to Blog