genai-code

Run Claude Code using Gemma 4, Gemini, or any Google GenAI model — completely local, no Anthropic API key needed.

Drop-in local proxy that lets you use Gemma 4, Gemini, or any Google GenAI model with Claude Code. No changes to your Claude Code workflow — just point it at the proxy.

Features

• Anthropic Messages API compatible — works transparently with Claude Code and any Anthropic SDK client
• Dynamic model passthrough — use any model you specify in the request (Gemma 4, Gemini, etc.)
• Streaming support — full SSE streaming with proper block indexing
• Tool / function calling — Claude Code's built-in tools (read, write, edit, bash) work via function calling translation
• Google Search grounding — optional web search via Google Search (enabled by default)
• Thinking config — control reasoning depth via thinking.budget_tokens
• Retry logic — automatic retries with exponential backoff on transient failures
• Connection handling — detects client disconnects mid-stream, avoids wasted API calls

Prerequisites

• Python 3.10+
• Google AI Studio API key — get one free

Quick Start

# 1. Clone the repo
git clone https://github.com/yourusername/genai-code.git
cd genai-code

# 2. Install dependencies
pip install -r requirements.txt

# 3. Set your API key
cp .env.example .env
# Edit .env and replace with your key: GEMINI_API_KEY=your_key_here

# 4. Start the proxy
uvicorn main:app --host 0.0.0.0 --port 8000

Verify it's running:

curl http://localhost:8000/health
# {"status":"ok","gemini_api_key_set":true,"model":"gemma-4-26b-a4b-it",...}

Use with Claude Code

Create a .claude/settings.json file in your project root (or global ~/.claude/settings.json):

{
    "$schema": "https://json.schemastore.org/claude-code-settings.json",
    "env": {
        "ANTHROPIC_AUTH_TOKEN": "sk-my-local-proxy-key",
        "ANTHROPIC_BASE_URL": "http://localhost:8000",
        "ANTHROPIC_MODEL": "gemma-4-26b-a4b-it",
        "ANTHROPIC_SMALL_FAST_MODEL": "gemma-4-26b-a4b-it"
    }
}

Then run claude normally — all requests route through Gemma 4.

Tip: Change ANTHROPIC_MODEL to any GenAI model (e.g., gemini-2.5-flash, gemma-4-29b-a4b-it). The proxy forwards the model name directly to the GenAI API.

API Reference

All endpoints are Anthropic Messages API compatible.

Method	Path	Description
GET	/	Status check
GET	/health	Health + API key status
GET	/v1/models	List available models
POST	/v1/messages	Chat endpoint (supports stream: true/false)
POST	/v1/messages/stream	Always-streaming alias
GET	/v1/tools	List proxy built-in tools

Request Format

{
    "model": "gemma-4-26b-a4b-it",
    "messages": [
        { "role": "user", "content": "Write a Python function to sort a list." }
    ],
    "max_tokens": 4096,
    "stream": false,
    "system": "You are an expert Python developer. Write clean, typed code.",
    "temperature": 0.2,

    "enable_google_search": true,
    "thinking": {
        "type": "enabled",
        "budget_tokens": 10000
    },
    "tools": [
        {
            "name": "get_weather",
            "description": "Get weather for a location",
            "input_schema": {
                "type": "object",
                "properties": {
                    "location": { "type": "string", "description": "City name" }
                },
                "required": ["location"]
            }
        }
    ]
}

Proxy-Specific Fields

Field	Type	Default	Description
model	string	gemma-4-26b-a4b-it	Passed through to GenAI — use any supported model
enable_google_search	bool	true	Enables Google Search grounding for factual queries
thinking	object	—	Maps budget_tokens to GenAI thinking level: <2k→LOW, 2k-8k→MEDIUM, ≥8k→HIGH
tools	array	—	Custom function-calling tools in Anthropic tool format

Note: When tools are provided, Google Search is automatically disabled to avoid SDK conflicts.

cURL Examples

Non-streaming:

curl http://localhost:8000/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: dummy" \
  -d '{
    "model": "gemma-4-26b-a4b-it",
    "messages": [{"role":"user","content":"What is 2+2?"}],
    "max_tokens": 256
  }'

Streaming:

curl http://localhost:8000/v1/messages \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-26b-a4b-it",
    "messages": [{"role":"user","content":"Tell me a joke"}],
    "stream": true
  }'

Configuration

Environment Variables

Variable	Required	Description
GEMINI_API_KEY	Yes	Your Google AI Studio API key

Set via .env file or export directly:

export GEMINI_API_KEY=your_key_here

Supported Models

Any model available through the Google GenAI SDK should work, including:

• gemma-4-26b-a4b-it (default)
• gemma-4-29b-a4b-it
• gemini-2.5-flash
• gemini-2.5-pro

Pass the model name in your request or via ANTHROPIC_MODEL in Claude Code settings.

Architecture

Claude Code ──POST /v1/messages──▶ FastAPI Proxy ──generateContent──▶ Google GenAI API
     ▲                                    │                                │
     │         SSE stream / JSON          │         Gemma 4 / Gemini       │
     └────────────────────────────────────└────────────────────────────────┘

1. Claude Code sends Anthropic-format requests to the proxy
2. Proxy converts messages/tools to GenAI Content and Tool formats
3. GenAI API response is converted back to Anthropic format (tool_use/text blocks)
4. Responses include proper anthropic-version headers and error shapes

Troubleshooting

Symptom	Likely Cause	Fix
GEMINI_API_KEY env var not set	Missing API key	Set GEMINI_API_KEY in .env or export it
502 error in Claude Code	GenAI API failure	Check logs for the exception type; retry logic handles most transient errors
Tools not being called	Model doesn't support function calling well	Try gemini-2.5-flash which has stronger tool-use capabilities
Empty responses	max_tokens too low or model error	Increase max_tokens (capped at 8192)
Slow responses	Thinking config or large context	Reduce budget_tokens or use a faster model like gemini-2.5-flash

Viewing Logs

The proxy logs to stdout. Set log level via uvicorn:

uvicorn main:app --host 0.0.0.0 --port 8000 --log-level debug

Contributing

Issues and pull requests welcome. Areas that could use help:

• Better tool schema translation edge cases (anyOf, oneOf)
• Token usage extraction from GenAI responses
• Multi-model /v1/models listing
• Docker image for easier deployment

License

MIT