If you want to use Claude Code with OpenRouter models like Qwen3 Coder or Kimi-K2 Pro this guide walks you through the setup.

Prerequisites

• Node.js (v18+)

• A working terminal or shell (e.g., bash, zsh)

• An OpenRouter API key (from https://openrouter.ai)

  Note that you must have bought at least $10 credits to be able to use even the :free quoted model.

Find the entire configuration on https://claudecoderouter.com/, although I must admit, I’ve spent close to a week going around circles trying to work it out, which nearly made me want to tear off my face 😵.

That’s when I came across AI Oriented Dev's youtube video, shoutout to him 🙌.

I discovered one key difference in how he approached it:

Before using Claude for coding help, it is to make sure to start the local router service:

claude start

Here is the complete installation process:

1. Install Claude Code Router

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

2. Configure the config.json file which needs to be located at ~/.claude-code-router. You are required to create it, if not present already.

   Here is the exact config.json file I’m using:

{
  "Providers": [
    {
      "name": "openrouter",
      "api_base_url": "https://openrouter.ai/api/v1/chat/completions",
      "api_key": "sk-openrouter-api-key",
      "models": [
        "anthropic/claude-3.5-sonnet",
        "google/gemini-2.5-pro-preview",
        "moonshotai/kimi-k2",
    "qwen/qwen3-coder:free"
      ],
      "transformer": {
        "use": ["openrouter"]
      }
    }
  ],
  "Router": {
    "default": "openrouter,qwen/qwen3-coder:free",
    // "default": "openrouter,qwen/qwen3-coder",
    // "default": "openrouter,moonshotai/kimi-k2",
    // "default": "openrouter,moonshotai/kimi-k2:free",

    "background": "openrouter,qwen/qwen3-coder:free",
    "think": "openrouter,qwen/qwen3-coder:free",
    "longContext": "openrouter,qwen/qwen3-coder:free"
  },
  "API_TIMEOUT_MS": 600000,
  "LOG": true
}

You might want to add the additionals fields:

API_TIMEOUT_MS=60000, which sets the timeout duration to 60,000ms, or 10 minutes for API requests. It is particularly useful when using slow or large models, which may take longer for complex completions.

LOG=true, which enables logging.

3. Start the local router service. This is the step that made all the difference, and it’s somehow now mentioned in the documentation as of this moments

claude start

4. Once the router is up and running, you can jump into coding mode:

claude code

And voila, it's working as expected!

That’s it from me! Let me know if you run into any issues- I’ve probably hit the same ones 😄

Cheers!

That’s where LiteLLM comes in. It acts as a local proxy that can translate Claude Code’s Anthropic-style requests into whatever OpenRouter (or any other provider) expects.

In this post, I’ll walk through how I got Claude Code to work with OpenRouter’s qwen/qwen3-coder: free model using LiteLLM along with all the missteps I made, and what finally worked.

What We’re Setting Up

• Claude Code CLI installed locally

• LiteLLM proxy server running on localhost

• Requests from Claude CLI routed to OpenRouter

• Fake Claude model names (like claude-3.5-sonnet) mapped to real OpenRouter models behind the scenes

Step 1: Install Dependencies

We’ll LiteLLM as a local proxy.

Install it via pip:

python3 -m venv LiteLLM
pip3 install "litellm[proxy]"

Step 2: Set Environment Variables

Set your secrets and endpoint overrides in your shell config (~/.zshrc or ~/.bashrc):

export LITELLM_MASTER_KEY=sk-1234      # This can be anything
export OPENROUTER_API_KEY=sk-your-openrouter-key

export ANTHROPIC_BASE_URL=http://localhost:4000
export ANTHROPIC_AUTH_TOKEN=$LITELLM_MASTER_KEY

Then reload your shell config:

source ~/.zshrc

Step 3: Create Your LiteLLM Config

Save the following YAML into config.yaml. This is what tells LiteLLM to act like it’s Anthropic, but actually forward requests to OpenRouter’s Qwen model.

model_list:
  - model_name: "claude-3.5-sonnet"
    litellm_params:
      model: "openrouter/qwen/qwen3-coder:free"
      model_provider: openrouter
      api_key: os.environ/OPENROUTER_API_KEY
      api_base: https://openrouter.ai/api/v1

litellm_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  database_type: none

Note that LiteLLM tends to be annoyingly pushy to make us use database. Hence, its important to add this field database_type: none .

Also, the trick here is in model_name. Claude Code thinks it’s calling claude-3.5-sonnet, but our proxy maps that name to Qwen.

Step 4: Start the LiteLLM Proxy

Run the proxy server:

litellm --config config.yaml

To make sure it’s working:

curl http://localhost:4000/health \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY"

You should see a healthy status with your model name listed.

You can also visit http://localhost:4000/health to ensure working.

Step 5: Use Claude CLI

Now that everything is in place, launch Claude with:

claude --model claude-3.5-sonnet

If everything went well, Claude will start, but it’ll actually be using OpenRouter’s Qwen model behind the scenes.

Mistakes I Made

1. The model name didn’t match

At one point, LiteLLM showed the model as /claude-3.5-sonnet instead of claude-3.5-sonnet. That happened because I accidentally added a leading slash. Claude CLI failed with a “model not found” error.

Fix: Make sure the model_name you enter in the config.yaml matches with the field you enter with the claude --model claude-3.5-sonnet

model_name: "claude-3.5-sonnet"

2. The wrong model ID for OpenRouter

I tried using "moonshot/kimi-k2:free" which OpenRouter didn’t support. It threw a model not found error.

Fix: Use curl https://openrouter.ai/api/v1/models with your API key to check which models are available and free.

3. Database errors

At one point I saw:

ModuleNotFoundError: No module named 'prisma'

This happened when I tried running features that require LiteLLM’s database integration, without setting up a DB.

Fix: Add this to your config to disable DB usage:

database_type: none

4. “No API key provided” even though I had one

This was just a case of forgetting to run:

source ~/.zshrc

after setting environment variables. Don’t skip that step.

Wrapping Up

Now I can run Claude CLI locally, while routing it through any OpenRouter-supported model, all while keeping the Anthropic API format that Claude Code expects.

This setup is super flexible. You could point Claude to Mistral, Moonshot, or even your own LLM behind a custom endpoint as long as LiteLLM can proxy to it.

There go 4 hours of my life I’ll neve get back :)