local-openai2anthropic 0.3.2

A lightweight proxy server that converts Anthropic Messages API to OpenAI API

local-openai2anthropic

English | 中文

A lightweight proxy that lets applications built with Claude SDK talk to locally-hosted OpenAI-compatible LLMs.

---

What Problem This Solves

Many local LLM tools (vLLM, SGLang, etc.) provide an OpenAI-compatible API. But if you've built your app using Anthropic's Claude SDK, you can't use them directly.

This proxy translates Claude SDK calls to OpenAI API format in real-time, enabling:

• Local LLM inference with Claude-based apps
• Offline development without cloud API costs
• Privacy-first AI - data never leaves your machine
• Seamless model switching between cloud and local
• Web Search tool - built-in Tavily web search for local models

---

Supported Local Backends

Currently tested and supported:

Backend	Description	Status
vLLM	High-throughput LLM inference	✅ Fully supported
SGLang	Fast structured language model serving	✅ Fully supported

Other OpenAI-compatible backends may work but are not fully tested.

---

Quick Start

1. Install

pip install local-openai2anthropic

2. Configure Your LLM Backend (Optional)

Option A: Start a local LLM server

If you don't have an LLM server running, you can start one locally:

Example with vLLM:

vllm serve meta-llama/Llama-2-7b-chat-hf
# vLLM starts OpenAI-compatible API at http://localhost:8000/v1

Or with SGLang:

sglang launch --model-path meta-llama/Llama-2-7b-chat-hf --port 8000
# SGLang starts at http://localhost:8000/v1

Option B: Use an existing OpenAI-compatible API

If you already have a deployed OpenAI-compatible API (local or remote), you can use it directly. Just note the base URL for the next step.

Examples:

• Local vLLM/SGLang: http://localhost:8000/v1
• Remote API: https://api.example.com/v1

Note: If you're using Ollama, it natively supports the Anthropic API format, so you don't need this proxy. Just point your Claude SDK directly to http://localhost:11434/v1.

3. Start the Proxy

Option A: Run in background (recommended)

export OA2A_OPENAI_BASE_URL=http://localhost:8000/v1  # Your local LLM endpoint
export OA2A_OPENAI_API_KEY=dummy  # Any value, not used by local backends

oa2a start              # Start server in background
# Server starts at http://localhost:8080

# View logs
oa2a logs               # Show last 50 lines of logs
oa2a logs -f            # Follow logs in real-time (Ctrl+C to exit)

# Check status
oa2a status             # Check if server is running

# Stop server
oa2a stop               # Stop background server

# Restart server
oa2a restart            # Restart with same settings

Option B: Run in foreground

export OA2A_OPENAI_BASE_URL=http://localhost:8000/v1
export OA2A_OPENAI_API_KEY=dummy

oa2a                    # Run server in foreground (blocking)
# Press Ctrl+C to stop

4. Use in Your App

import anthropic

client = anthropic.Anthropic(
    base_url="http://localhost:8080",  # Point to proxy
    api_key="dummy-key",  # Not used
)

message = client.messages.create(
    model="meta-llama/Llama-2-7b-chat-hf",  # Your local model name
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)

print(message.content[0].text)

---

Using with Claude Code

You can configure Claude Code to use your local LLM through this proxy.

Configuration Steps

1. Edit Claude Code config file at ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:8080",
    "ANTHROPIC_API_KEY": "dummy-key",
    "ANTHROPIC_MODEL": "meta-llama/Llama-2-7b-chat-hf",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "meta-llama/Llama-2-7b-chat-hf",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "meta-llama/Llama-2-7b-chat-hf",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "meta-llama/Llama-2-7b-chat-hf",
    "ANTHROPIC_REASONING_MODEL": "meta-llama/Llama-2-7b-chat-hf"
  }
}

Variable	Description
ANTHROPIC_MODEL	General model setting
ANTHROPIC_DEFAULT_SONNET_MODEL	Default model for Sonnet mode (Claude Code default)
ANTHROPIC_DEFAULT_OPUS_MODEL	Default model for Opus mode
ANTHROPIC_DEFAULT_HAIKU_MODEL	Default model for Haiku mode
ANTHROPIC_REASONING_MODEL	Default model for reasoning tasks

2. Or set environment variables before running Claude Code:

export ANTHROPIC_BASE_URL=http://localhost:8080
export ANTHROPIC_API_KEY=dummy-key

claude

Complete Workflow Example

Make sure ~/.claude/settings.json is configured as described above.

Terminal 1 - Start your local LLM:

vllm serve meta-llama/Llama-2-7b-chat-hf

Terminal 2 - Start the proxy (background mode):

export OA2A_OPENAI_BASE_URL=http://localhost:8000/v1
export OA2A_OPENAI_API_KEY=dummy
export OA2A_TAVILY_API_KEY="tvly-your-tavily-api-key"  # Optional: enable web search

oa2a start

Terminal 3 - Launch Claude Code:

claude

Now Claude Code will use your local LLM instead of the cloud API.

To stop the proxy:

oa2a stop

---

Features

• ✅ Streaming responses - Real-time token streaming via SSE
• ✅ Tool calling - Local LLM function calling support
• ✅ Vision models - Multi-modal input for vision-capable models
• ✅ Web Search - Give your local LLM internet access (see below)
• ✅ Thinking mode - Supports reasoning/thinking model outputs

---

Web Search Capability 🔍

Bridge the gap: Give your local LLM the web search power that Claude Code users enjoy!

When using locally-hosted models with Claude Code, you lose access to the built-in web search tool. This proxy fills that gap by providing a server-side web search implementation powered by Tavily.

The Problem

Scenario	Web Search Available?
Using Claude (cloud) in Claude Code	✅ Built-in
Using local vLLM/SGLang in Claude Code	❌ Not available
Using this proxy + local LLM	✅ Enabled via Tavily

How It Works

Claude Code → Anthropic SDK → This Proxy → Local LLM
                                      ↓
                                 Tavily API (Web Search)

The proxy intercepts web_search_20250305 tool calls and handles them directly, regardless of whether your local model supports web search natively.

Setup Tavily Search

1. Get a free API key at tavily.com - generous free tier available

2. Configure the proxy:

export OA2A_OPENAI_BASE_URL=http://localhost:8000/v1
export OA2A_OPENAI_API_KEY=dummy
export OA2A_TAVILY_API_KEY="tvly-your-tavily-api-key"  # Enable web search

oa2a

3. Use in your app:

import anthropic

client = anthropic.Anthropic(
    base_url="http://localhost:8080",
    api_key="dummy-key",
)

message = client.messages.create(
    model="meta-llama/Llama-2-7b-chat-hf",
    max_tokens=1024,
    tools=[
        {
            "name": "web_search_20250305",
            "description": "Search the web for current information",
            "input_schema": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "Search query"},
                },
                "required": ["query"],
            },
        }
    ],
    messages=[{"role": "user", "content": "What happened in AI today?"}],
)

if message.stop_reason == "tool_use":
    tool_use = message.content[-1]
    print(f"Searching: {tool_use.input}")
    # The proxy automatically calls Tavily and returns results

Tavily Configuration Options

Variable	Default	Description
OA2A_TAVILY_API_KEY	-	Your Tavily API key (get free at tavily.com)
OA2A_TAVILY_MAX_RESULTS	5	Number of search results to return
OA2A_TAVILY_TIMEOUT	30	Search timeout in seconds
OA2A_WEBSEARCH_MAX_USES	5	Max search calls per request

---

Configuration

Variable	Required	Default	Description
OA2A_OPENAI_BASE_URL	✅	-	Your local LLM's OpenAI-compatible endpoint
OA2A_OPENAI_API_KEY	✅	-	Any value (local backends usually ignore this)
OA2A_PORT	❌	8080	Proxy server port
OA2A_HOST	❌	0.0.0.0	Proxy server host
OA2A_TAVILY_API_KEY	❌	-	Enable web search (tavily.com)

---

Architecture

Your App (Claude SDK)
         │
         ▼
┌─────────────────────┐
│  local-openai2anthropic  │  ← This proxy
│  (Port 8080)        │
└─────────────────────┘
         │
         ▼
Your Local LLM Server
(vLLM / SGLang)
(OpenAI-compatible API)

---

Development

git clone https://github.com/dongfangzan/local-openai2anthropic.git
cd local-openai2anthropic
pip install -e ".[dev]"

pytest

License

Apache License 2.0