Claude Max API Proxy

Turn your $200/month Claude Max subscription into a full OpenAI-compatible API — no extra cost.

A lightweight local proxy that wraps the Claude Code CLI as an OpenAI-compatible API server. Any app that speaks the OpenAI format can use your Claude Max subscription directly.

---

Why?

• You already pay for Claude Max ($200/month) and want to use it everywhere
• You use OpenClaw, Continue.dev, or other tools that need an API endpoint
• You want to use Claude in scripts and workflows without per-message API costs

Anthropic blocks third-party apps from using your Max subscription's OAuth tokens directly. But the Claude Code CLI is allowed to use them. This proxy bridges that gap.

---

Quick Start

Prerequisites

• Node.js 20+ — nodejs.org
• Claude Code CLI — installed and authenticated:

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

• Claude Max subscription ($200/month plan from Anthropic)

Install & Run

git clone https://github.com/mattschwen/claude-max-api-proxy.git
cd claude-max-api-proxy
npm install
npm start

The server starts on http://localhost:3456. Keep the terminal open.

Verify

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

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

# 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": "Tell me a joke"}], "stream": true}'

---

Connect Your Apps

Point any OpenAI-compatible app at the proxy:

Setting	Value
API Base URL	http://localhost:3456/v1
API Key	Any string (the proxy ignores it)
Model	claude-opus-4, claude-sonnet-4, or claude-haiku-4

OpenClaw

Set the API endpoint in OpenClaw's settings to http://localhost:3456/v1 with any API key.

Continue.dev (VS Code)

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

Python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:3456/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="claude-sonnet-4",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

---

Available Models

Model ID	Family	Best For
claude-opus-4-6 / claude-opus-4	Opus	Complex reasoning, analysis, coding
claude-sonnet-4-6 / claude-sonnet-4	Sonnet	General use, balanced speed and capability
claude-haiku-4-5 / claude-haiku-4	Haiku	Quick questions, simple tasks

You can also use bare aliases: opus, sonnet, haiku.

---

Reliability Features

The proxy is built for long-running, production-like use with aggressive reliability measures:

Activity-Based Stall Detection

Each model has an activity timeout that resets on every content token. If a subprocess stops producing output, it's killed and the queue unblocks — no more 30-minute hangs.

Model	Activity Timeout	Hard Timeout
Opus	120s	30min
Sonnet	60s	10min
Haiku	30s	2min

Kill Escalation

Stuck processes that ignore SIGTERM are escalated to SIGKILL after a 5-second grace period, preventing zombie processes.

Per-Conversation Queue

Requests for the same conversation are serialized (FIFO). Different conversations run in parallel. Each queue slot has an absolute timeout to prevent permanent blocking.

Session Resume with Failure Tracking

The proxy automatically resumes Claude CLI sessions for multi-turn conversations. If resume fails twice consecutively, the session is invalidated and a fresh one is created — no silent stalls.

Retry Logic

On subprocess failure, the proxy retries once with a fresh session (no resume). Each attempt gets its own clean event handler wiring.

Graceful Shutdown

On SIGTERM/SIGINT: stops accepting new connections, waits up to 30s for in-flight requests, kills remaining subprocesses, saves session state, then exits cleanly.

Structured Logging

All key events emit structured JSON logs: subprocess.spawn, subprocess.stall, subprocess.kill, request.complete, session.invalidate, etc.

---

API Reference

Endpoint	Method	Description
/health	GET	Health check with subprocess counts, queue status, stall stats, session failure rates
/v1/models	GET	List available Claude models
/v1/chat/completions	POST	Chat completions (streaming and non-streaming)

Extended Thinking

Pass the thinking parameter to enable extended thinking mode:

{
  "model": "claude-opus-4",
  "messages": [{"role": "user", "content": "Solve this complex problem..."}],
  "thinking": {"type": "enabled", "budget_tokens": 10000},
  "stream": true
}

Timeouts are automatically tripled when extended thinking is active.

Conversation Continuity

Set the user field to a stable conversation ID to maintain context across requests:

{
  "model": "claude-sonnet-4",
  "user": "my-conversation-123",
  "messages": [{"role": "user", "content": "Follow-up question..."}]
}

The proxy uses Claude CLI's --resume flag to continue the conversation without replaying history.

---

Architecture

Your App (OpenClaw, Continue.dev, Python, etc.)
    |
    v
  HTTP Request (OpenAI format)
    |
    v
  This Proxy (localhost:3456)
    |
    v
  Claude Code CLI (subprocess per request)
    |
    v
  Claude Max subscription (via OAuth)
    |
    v
  Anthropic's API
    |
    v
  Response streams back to your app

Key internals:

• Express server with SSE streaming and keepalive
• Per-conversation FIFO queue for request serialization
• Subprocess manager with structured event parsing
• Session manager with disk persistence and TTL
• SQLite conversation store for message history and metrics
• Warm-up pool that pre-heats the OS process cache

---

Auto-Start on macOS

See docs/macos-setup.md for LaunchAgent configuration.

---

Development

# Build from TypeScript sources
npm run build

# Watch mode
npm run dev

# Run
npm start

# Custom port
node dist/server/standalone.js 8080

The TypeScript sources are in src/, compiled output in dist/.

---

Troubleshooting

"Claude CLI not found"

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

Port 3456 already in use

# Find what's using it
lsof -i :3456

# Use a different port
node dist/server/standalone.js 3457

Subprocess hangs / no response

Check the structured logs — stall detection should auto-kill hung processes. If issues persist, check /health for active subprocess PIDs and queue status.

Streaming gives no output

Use the -N flag with curl to disable buffering:

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

---

Security

• No API keys stored — authentication handled by Claude CLI's secure keychain
• Localhost only — binds to 127.0.0.1 by default
• No shell execution — all subprocesses spawned with spawn(), not exec()
• No secrets in source — environment variables cleaned before subprocess spawn

---

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.