OpenClaw Subscription Billing Proxy

Route your OpenClaw API requests through your Claude Max/Pro subscription instead of Extra Usage billing.

What This Does

After Anthropic revoked subscription billing for third-party tools (April 4, 2026), OpenClaw requests are billed to Extra Usage. This proxy sits between OpenClaw and the Anthropic API, injecting Claude Code's billing identifier so requests use your existing subscription.

Zero cost increase. Full OpenClaw functionality. No code changes to OpenClaw.

How It Works

The proxy performs bidirectional request/response processing:

Outbound (request to API):

1. Billing Header -- Injects an 84-character Claude Code billing identifier into the system prompt
2. Token Swap -- Replaces OpenClaw's auth token with your Claude Code OAuth token
3. Sanitization -- Replaces trigger phrases that Anthropic's streaming classifier detects (8 default patterns)

Inbound (response to OpenClaw): 4. Reverse Mapping -- Restores original tool names, file paths, and identifiers so OpenClaw sees its real terms

This ensures Anthropic sees a "Claude Code" request while OpenClaw sees its original tool names and paths.

Requirements

• Node.js 18+
• Claude Max or Pro subscription
• Claude Code CLI installed and authenticated
• OpenClaw installed and running

Installing Claude Code CLI (if not already installed)

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

Then authenticate with your Claude account:

claude auth login

This opens a browser window to sign in with your Claude Max/Pro account. Once authenticated, credentials are stored at ~/.claude/.credentials.json. The proxy reads from this file.

Verify it worked:

claude auth status
# Should show: loggedIn: true, subscriptionType: max (or pro)

Quick Start

# 1. Clone
git clone https://github.com/zacdcook/openclaw-billing-proxy
cd openclaw-billing-proxy

# 2. Run setup (auto-detects your config)
node setup.js

# 3. Start the proxy
node proxy.js

# 4. Update OpenClaw config
# In ~/.openclaw/openclaw.json, change:
#   "baseUrl": "https://api.anthropic.com"
# to:
#   "baseUrl": "http://127.0.0.1:18801"

# 5. Restart your OpenClaw gateway

Configuration

The config.json file (generated by setup or created manually):

{
  "port": 18801,
  "credentialsPath": "~/.claude/.credentials.json",
  "replacements": [
    ["OpenClaw", "OCPlatform"],
    ["openclaw", "ocplatform"],
    ["sessions_spawn", "create_task"],
    ["sessions_list", "list_tasks"],
    ["sessions_history", "get_history"],
    ["sessions_send", "send_to_task"],
    ["sessions_yield", "yield_task"],
    ["running inside", "running on"]
  ],
  "reverseMap": [
    ["OCPlatform", "OpenClaw"],
    ["ocplatform", "openclaw"],
    ["create_task", "sessions_spawn"],
    ["list_tasks", "sessions_list"],
    ["get_history", "sessions_history"],
    ["send_to_task", "sessions_send"],
    ["yield_task", "sessions_yield"]
  ]
}

Replacement Rules

replacements -- Applied to outbound requests. Each [find, replace] pair does a raw string replacement on the request body before forwarding to the API.

reverseMap -- Applied to inbound responses. Each [sanitized, original] pair restores terms back to their original form before returning to OpenClaw. This ensures tool names, file paths, and identifiers work correctly.

Important: Use space-free replacements for lowercase openclaw. Replacing .openclaw/ with .assistant platform/ (with a space) breaks filesystem paths in tool calls. Use ocplatform or similar instead.

Adding New Patterns

If your OpenClaw version has additional sessions_* tools (they add new ones across versions), add them to both replacements and reverseMap:

{
  "replacements": [
    ["sessions_new_tool", "new_task_tool"]
  ],
  "reverseMap": [
    ["new_task_tool", "sessions_new_tool"]
  ]
}

If you have a custom assistant name that Anthropic blocks (test by checking if requests fail with the name present but pass without it), add it the same way.

Running as a Service

Linux (systemd)

sudo tee /etc/systemd/system/openclaw-proxy.service << EOF
[Unit]
Description=OpenClaw Billing Proxy
After=network.target

[Service]
ExecStart=/usr/bin/node /path/to/proxy.js
Restart=always
User=YOUR_USER

[Install]
WantedBy=multi-user.target
EOF

sudo systemctl enable openclaw-proxy
sudo systemctl start openclaw-proxy

Windows (startup)

Add to your gateway.cmd before the gateway launch:

start "Billing Proxy" /min node "C:\path\to\proxy.js"
timeout /t 2 /nobreak >nul

PM2

pm2 start proxy.js --name openclaw-proxy
pm2 save

Token Refresh

Claude Code's OAuth token expires every ~24 hours. The proxy reads the token fresh from disk on each request. To refresh:

• Easiest: Open Claude Code CLI briefly -- it auto-refreshes on startup
• Automated: Set up a cron that runs claude -p "ping" --max-turns 1 --no-session-persistence daily (triggers auth refresh)

Health Check

curl http://127.0.0.1:18801/health

Returns token status, request count, uptime, subscription type, and pattern counts.

How Anthropic's Detection Works

Anthropic uses two mechanisms to detect third-party tools:

1. Billing Header -- The API checks the system prompt for x-anthropic-billing-header. Without it, OAuth requests go to Extra Usage. This is a simple string match on 84 characters injected by Claude Code's SDK.

2. Streaming Classifier -- Anthropic's classifier scans the request for specific trigger phrases. The verified triggers are:

   • OpenClaw (the platform name -- case insensitive)
   • sessions_spawn, sessions_list, sessions_history, sessions_send, sessions_yield (OpenClaw session management tools -- new tools are added across versions)
   • running inside + platform name (the self-declaration phrase OpenClaw injects)

   If detected, the response is refused with stop_reason: "refusal" or a billing error.

What is NOT detected: Assistant names, workspace filenames (AGENTS.md, SOUL.md), config paths (.openclaw/), plugin names (lossless-claw), individual non-session tool names (exec, lcm_grep, gateway, cron, etc.), bot names, or runtime references. Only the patterns above trigger rejection.

Why Reverse Mapping Matters

Without reverse mapping, the model sees sanitized paths in its context (e.g., .ocplatform/workspace/scripts/) and uses them for tool calls. But the actual filesystem has .openclaw/. The reverse mapping translates API responses back to original terms before OpenClaw processes them, ensuring:

• Tool names match OpenClaw's tool registry
• File paths match the actual filesystem
• Session management commands use correct identifiers

Rollback

Change baseUrl back to https://api.anthropic.com in openclaw.json and restart the gateway. Enable Extra Usage in your Claude settings if needed.

Disclaimer

This is an unofficial workaround. Anthropic may change their detection at any time. Use at your own risk. This proxy does not modify OpenClaw or Claude Code -- it's a transparent HTTP middleman.

License

MIT