Using Alternate Models with Claude Code via Claude Code Router

Why Use Alternate Models?

Claude Code is fantastic, but the default models can get expensive when you are running long sessions. Claude Code Router (CCR) is a local proxy that translates Anthropic's API protocol to other providers, letting you use models from Gemini, DeepSeek, OpenAI, and others directly inside Claude Code. You get the full Claude Code experience (slash commands, tools, MCP, hooks) backed by cheaper models of your choice.

Setup

Installing CCR

CCR uses SQLite for its config, which means the better-sqlite3 native module needs to compile during install. You need to allow scripts for that package:

npm install -g --allow-scripts=better-sqlite3 @musistudio/claude-code-router

Configuring DeepSeek

You will need a DeepSeek account. Sign up, grab your API key from the dashboard, and check out the pricing page. DeepSeek's API pricing is significantly cheaper than Anthropic's, especially for the flash model.

Start CCR and open the admin UI:

ccr start
ccr ui

In the UI, configure three things:

1. Provider: add the DeepSeek provider with your API key
2. Models: add the models you want to use (deepseek-v4-pro for heavy lifting, deepseek-v4-flash for the small/fast model)
3. Agent: set default models for your Claude Code agent

CCR automatically writes the required config to ~/.claude/settings.json: the apiKeyHelper path, model environment variables, and the local proxy base URL.

Once configured, launch Claude Code through the router:

ccr 'Claude Code'

Claude Code will now use DeepSeek models through the local proxy without noticing anything is different.

The Gemini Dead End

I initially tried using Gemini 3.5 Flash (gemini/gemini-3.5-flash) as my model. Single-turn requests worked perfectly:

curl -s "http://127.0.0.1:3456/v1/messages" \
  -H "Content-Type: application/json" \
  -H "x-api-key: ccr-profile-..." \
  -H "anthropic-version: 2023-06-01" \
  -d '{"model":"gemini/gemini-3.5-flash","max_tokens":500,
       "messages":[{"role":"user","content":"Who made you?"}]}'

That returned a clean response without issues. But Claude Code broke immediately with an API Error: 400 All target providers failed on anything involving tool use, which is pretty much everything Claude Code does (skills, file reads, bash commands).

The Root Cause

The problem is Google's Gemini 3 API requirement for a thought_signature on function call parts. When Claude Code sends a tool call, CCR translates it from Anthropic's tool_use format to Gemini's functionCall format. On the follow-up turn, when CCR sends the tool result back, Gemini rejects the request because the thought_signature is missing. Anthropic's API has no equivalent field, so there is nothing for CCR to pass through.

I reproduced the exact failure with a minimal curl that simulated a tool-use conversation:

curl -sv http://127.0.0.1:3456/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: ccr-profile-..." \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model":"gemini/gemini-3.5-flash",
    "max_tokens":1000,
    "tools":[{"name":"get_weather",
      "description":"Get the weather for a location",
      "input_schema":{"type":"object",
        "properties":{"location":{"type":"string"}},
        "required":["location"]}}],
    "messages":[
      {"role":"user",
       "content":"What is the weather in Paris?"},
      {"role":"assistant",
       "content":[{"type":"tool_use","id":"tool_1",
         "name":"get_weather",
         "input":{"location":"Paris"}}]},
      {"role":"user",
       "content":[{"type":"tool_result",
         "tool_use_id":"tool_1",
         "content":"18°C, partly cloudy"}]}
    ]
  }'

The response confirmed it:

Function call is missing a thought_signature in functionCall parts.
This is required for tools to work correctly, and missing
thought_signature may lead to degraded model performance.

This is a known issue in CCR. Gemini 2.x models (gemini/gemini-2.5-flash or gemini/gemini-2.5-pro) do not have the thought_signature requirement and should work fine, but I wanted to try something different.

Switching to DeepSeek

I switched to the DeepSeek provider with two models:

• deepseek-v4-pro: for the main Claude Code model, handling code generation, reasoning and complex tasks
• deepseek-v4-flash: for the small/fast model, handling quick completions, light edits and background tasks

The switch was seamless. No thought_signature issues, no protocol translation bugs, no mysterious 400 errors. Tool use works correctly across multi-turn conversations, which means skills, file operations, bash execution and MCP tools all function as expected.

Pricing is also reasonable. DeepSeek charges per million tokens and both the flash and pro models are considerably cheaper than Anthropic's equivalents. For long Claude Code sessions that rack up hundreds of thousands of tokens, the savings add up quickly.

Summary

Claude Code Router is a solid way to use cheaper models inside Claude Code. If you go the Gemini route, stick to 2.x models to avoid the thought_signature issue. I landed on DeepSeek with deepseek-v4-pro and deepseek-v4-flash and have been happy with the balance of cost and capability.