anthropic-bridge

A proxy server that exposes an Anthropic Messages API-compatible endpoint while routing requests to various LLM providers.

Features

• Anthropic Messages API compatible (/v1/messages)
• JSON responses by default, SSE when stream: true
• Tool/function calling support
• Multi-round conversations
• Support for multiple providers: OpenAI, GitHub Copilot, OpenRouter (Gemini, Grok, DeepSeek, Qwen, MiniMax, etc.)
• Extended thinking/reasoning support for compatible models
• Reasoning cache for Gemini models across tool call rounds
• Approximate /v1/messages/count_tokens support

Installation

pip install anthropic-bridge

For development:

git clone https://github.com/michaelgendy/anthropic-bridge.git
cd anthropic-bridge
pip install -e ".[test,dev]"

Usage

Start the bridge server:

anthropic-bridge --port 8080

All providers are configured via environment variables. The server is designed to run inside managed environments (e.g. claudex sandboxes) where tokens are injected automatically.

Provider Examples

from anthropic import Anthropic

client = Anthropic(
    api_key="not-used",
    base_url="http://localhost:8080"
)

# OpenAI (via ChatGPT subscription)
response = client.messages.create(
    model="openai/gpt-5.2",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}]
)

# GitHub Copilot
response = client.messages.create(
    model="copilot/gpt-5.3-codex",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}]
)

# OpenRouter
response = client.messages.create(
    model="openrouter/google/gemini-3-pro-preview",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}]
)

client.messages.create(...) works as a normal non-streaming Anthropic request. If you send stream=True, the bridge returns Anthropic-compatible SSE events.

Thinking/Reasoning

Use the thinking parameter to control reasoning effort (supported on OpenAI and compatible models):

response = client.messages.create(
    model="openai/gpt-5.2",
    max_tokens=1024,
    thinking={"budget_tokens": 15000},  # Maps to "high" effort
    messages=[{"role": "user", "content": "Solve this problem..."}]
)

Budget Tokens	Reasoning Effort
1 - 9,999	low
10,000 - 14,999	medium
15,000 - 31,999	high
32,000+	xhigh

API Endpoints

Endpoint	Method	Description
/	GET	Health check
/health	GET	Health check
/v1/messages	POST	Anthropic Messages API. Returns JSON unless stream: true is set.
/v1/messages/count_tokens	POST	Approximate token counting for structured Anthropic inputs.

Configuration

All providers are configured exclusively through environment variables:

Environment Variable	Required	Description
OPENROUTER_API_KEY	No	OpenRouter API key (required for openrouter/* models)
GITHUB_COPILOT_TOKEN	No	GitHub Copilot OAuth token (required for copilot/* models)

OpenAI models (openai/*) authenticate via the Codex CLI auth file (~/.codex/auth.json), which is set up externally.

CLI Flag	Default	Description
--port	8080	Port to run on
--host	127.0.0.1	Host to bind to

Model Routing

• openai/* → Direct OpenAI API (via Codex CLI auth)
• copilot/* → GitHub Copilot API (via GITHUB_COPILOT_TOKEN)
• openrouter/* → OpenRouter API (via OPENROUTER_API_KEY)
• Any other model → Falls back in this order: OpenRouter, Copilot, OpenAI

Explicit prefixes are strict. If you request openai/*, copilot/*, or openrouter/* and that backend is not configured, the bridge returns an authentication/configuration error instead of silently rerouting to another provider.

Architecture

• server.py owns provider selection, Anthropic route handling, SSE passthrough, and non-stream aggregation.
• providers/openrouter/client.py handles OpenRouter chat-completions streaming and provider-specific request tweaks.
• providers/openai/client.py and providers/copilot/client.py translate Anthropic requests onto Responses API or chat-completions style upstream APIs.
• transform.py converts Anthropic messages, tools, and tool choice into the upstream shapes used by the providers.

Provider Notes

This project does not keep a frozen catalog of upstream model IDs. Use the provider prefixes with the model IDs currently offered by that provider.

Provider-specific optimizations currently exist for:

• Google Gemini (openrouter/google/*) - Reasoning detail caching
• OpenAI (openrouter/openai/*) - Extended thinking support
• xAI Grok (openrouter/x-ai/*) - XML tool call parsing
• DeepSeek (openrouter/deepseek/*)
• Qwen (openrouter/qwen/*)
• MiniMax (openrouter/minimax/*)

Development

pytest tests/ -v runs the deterministic test suite by default. Live upstream tests are opt-in and require ANTHROPIC_BRIDGE_LIVE_TESTS=1.

License

MIT