Claude Max API Proxy

Actively maintained fork of atalovesyou/claude-max-api-proxy with OpenClaw integration, improved streaming, and expanded model support.

Use your Claude Max subscription ($200/month) with any OpenAI-compatible client — no separate API costs!

This proxy wraps the Claude Code CLI as a subprocess and exposes an OpenAI-compatible HTTP API, allowing tools like OpenClaw, Continue.dev, or any OpenAI-compatible client to use your Claude Max subscription instead of paying per-API-call.

Why This Exists

Approach	Cost	Limitation
Claude API	~$15/M input, ~$75/M output tokens	Pay per use
Claude Max	$200/month flat	OAuth blocked for third-party API use
This Proxy	$0 extra (uses Max subscription)	Routes through CLI

Anthropic blocks OAuth tokens from being used directly with third-party API clients. However, the Claude Code CLI can use OAuth tokens. This proxy bridges that gap by wrapping the CLI and exposing a standard API.

How It Works

Your App (OpenClaw, Continue.dev, etc.)
         ↓
    HTTP Request (OpenAI format)
         ↓
   Claude Max API Proxy (this project)
         ↓
   Claude Code CLI (subprocess)
         ↓
   OAuth Token (from Max subscription)
         ↓
   Anthropic API
         ↓
   Response → OpenAI format → Your App

Features

• OpenAI-compatible API — Works with any client that supports OpenAI's API format
• Streaming support — Real-time token streaming via Server-Sent Events
• Multiple models — Claude Opus, Sonnet, and Haiku with flexible model aliases
• OpenClaw integration — Automatic tool name mapping and system prompt adaptation
• Content block handling — Proper text block separators for multi-block responses
• Session management — Maintains conversation context via session IDs
• Auto-start service — Optional LaunchAgent for macOS
• Zero configuration — Uses existing Claude CLI authentication
• Secure by design — Uses spawn() to prevent shell injection

What's Different from the Original

• OpenClaw tool mapping — Maps OpenClaw tool names (exec, read, web_search, etc.) to Claude Code equivalents (Bash, Read, WebSearch)
• System prompt stripping — Removes OpenClaw-specific tooling sections that confuse the CLI
• Content block support — Handles input_text content blocks and multi-block text separators
• Tool call types — Full OpenAI tool call type definitions for streaming and non-streaming
• Improved streaming — Better SSE handling with connection confirmation and client disconnect detection

Prerequisites

1. Claude Max subscription ($200/month) — Subscribe here
2. Claude Code CLI installed and authenticated:

   npm install -g @anthropic-ai/claude-code
   claude auth login

Installation

# Clone the repository
git clone https://github.com/wende/claude-max-api-proxy.git
cd claude-max-api-proxy

# Install dependencies
npm install

# Build
npm run build

Usage

Start the server

npm start
# or
node dist/server/standalone.js

The server runs at http://localhost:3456 by default. Pass a custom port as an argument:

node dist/server/standalone.js 8080

Test it

# Health check
curl http://localhost:3456/health

# List models
curl http://localhost:3456/v1/models

# Chat completion (non-streaming)
curl -X POST http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

# Chat completion (streaming)
curl -N -X POST http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

API Endpoints

Endpoint	Method	Description
/health	GET	Health check
/v1/models	GET	List available models
/v1/chat/completions	POST	Chat completions (streaming & non-streaming)

Available Models

Model ID	Alias	CLI Model
claude-opus-4	opus	Claude Opus
claude-sonnet-4	sonnet	Claude Sonnet
claude-haiku-4	haiku	Claude Haiku

All model IDs also accept a claude-code-cli/ prefix (e.g., claude-code-cli/claude-opus-4). Unknown models default to Opus.

Configuration with Popular Tools

OpenClaw

OpenClaw works with this proxy out of the box. The proxy automatically maps OpenClaw tool names to Claude Code equivalents and strips conflicting tooling sections from system prompts.

Continue.dev

Add to your Continue config:

{
  "models": [{
    "title": "Claude (Max)",
    "provider": "openai",
    "model": "claude-sonnet-4",
    "apiBase": "http://localhost:3456/v1",
    "apiKey": "not-needed"
  }]
}

Generic OpenAI Client (Python)

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:3456/v1",
    api_key="not-needed"  # Any value works
)

response = client.chat.completions.create(
    model="claude-sonnet-4",
    messages=[{"role": "user", "content": "Hello!"}]
)

Auto-Start on macOS

The proxy can run as a macOS LaunchAgent on port 3456.

Plist location: ~/Library/LaunchAgents/com.openclaw.claude-max-proxy.plist

# Start the service
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.openclaw.claude-max-proxy.plist

# Restart
launchctl kickstart -k gui/$(id -u)/com.openclaw.claude-max-proxy

# Stop
launchctl bootout gui/$(id -u)/com.openclaw.claude-max-proxy

# Check status
launchctl list com.openclaw.claude-max-proxy

Architecture

src/
├── types/
│   ├── claude-cli.ts      # Claude CLI JSON streaming types + type guards
│   └── openai.ts          # OpenAI API types (including tool calls)
├── adapter/
│   ├── openai-to-cli.ts   # Convert OpenAI requests → CLI format
│   └── cli-to-openai.ts   # Convert CLI responses → OpenAI format
├── subprocess/
│   └── manager.ts         # Claude CLI subprocess + OpenClaw tool mapping
├── session/
│   └── manager.ts         # Session ID mapping
├── server/
│   ├── index.ts           # Express server setup
│   ├── routes.ts          # API route handlers
│   └── standalone.ts      # Entry point
└── index.ts               # Package exports

Security

• Uses Node.js spawn() instead of shell execution to prevent injection attacks
• No API keys stored or transmitted by this proxy
• All authentication handled by Claude CLI's secure keychain storage
• Prompts passed as CLI arguments, not through shell interpretation

Troubleshooting

"Claude CLI not found"

Install and authenticate the CLI:

npm install -g @anthropic-ai/claude-code
claude auth login

Streaming returns immediately with no content

Ensure you're using -N flag with curl (disables buffering):

curl -N -X POST http://localhost:3456/v1/chat/completions ...

Server won't start

Check that the Claude CLI is in your PATH:

which claude

Contributing

Contributions welcome! Please submit PRs with tests.

License

MIT

Acknowledgments

• Originally created by atalovesyou
• Built for use with OpenClaw
• Powered by Claude Code CLI