@authloop-ai/mcp

MCP server for AuthLoop — human-in-the-loop authentication for AI agents.

Exposes the authloop_handoff tool so AI agents can hand off auth challenges (OTP, captcha, password) to a human via the Model Context Protocol.

Setup

Add to your MCP client config:

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "authloop": {
      "command": "npx",
      "args": ["-y", "@authloop-ai/mcp"],
      "env": { "AUTHLOOP_API_KEY": "al_live_..." }
    }
  }
}

OpenClaw (openclaw.json):

{
  "mcp": {
    "servers": {
      "authloop": {
        "command": "npx",
        "args": ["-y", "@authloop-ai/mcp"],
        "env": { "AUTHLOOP_API_KEY": "al_live_..." }
      }
    }
  }
}

Tool: authloop_handoff

Hand off a login or auth challenge to a human who can resolve it remotely.

Input

Output

{
  "session_url": "https://authloop.ai/session/sess_...",
  "status": "resolved"
}

Status is one of: resolved, cancelled, error, timeout.

How It Works

Agent                    MCP Server                  Relay                    Human
  │                          │                         │                       │
  │ calls authloop_handoff   │                         │                       │
  │─────────────────────────→│                         │                       │
  │                          │ POST /session            │                       │
  │                          │────────────────────────→ │                       │
  │                          │ connect WSS (instant)    │                       │
  │                          │────────────────────────→ │                       │
  │                          │                          │  viewer_connected     │
  │                          │ ←────────────────────── │ ←─────────────────    │
  │                          │                          │                       │
  │                          │ CDP screencast ──→ JPEG frames over WSS ──→    │
  │                          │                          │                       │
  │                          │     E2EE encrypted input over WSS ←─────────   │
  │                          │ CDP dispatch ←────       │                       │
  │                          │                          │                       │
  │                          │        { "type": "resolved" } ←────────────    │
  │                          │ POST /session/:id/resolve│                       │
  │  { status: "resolved" }  │                         │                       │
  │←─────────────────────────│                         │                       │

1. Agent calls authloop_handoff when it hits an auth wall
2. MCP server creates a session via the AuthLoop API
3. Agent sends the session_url to the human (Telegram, Slack, etc.)
4. MCP connects to the relay WebSocket immediately (no polling)
5. Human opens the URL — relay notifies MCP instantly via viewer_connected
6. MCP starts CDP screencast, JPEG frames stream to the human's browser
7. Human sees the live browser, clicks and types
8. All input is end-to-end encrypted (E2EE) and dispatched to the browser via CDP
9. Human clicks "Done" to resolve or "Cancel" to abort — agent continues with the result

Security

End-to-End Encryption (E2EE)

All user input is end-to-end encrypted between the human's browser and the MCP server. The relay server cannot read what the human types or clicks.

• Key exchange: ECDH on P-256 curve
• Encryption: AES-256-GCM (12-byte IV, 16-byte auth tag)
• What's encrypted: all user input — keystrokes, clicks, scroll, paste, navigation, resolve/cancel
• What's NOT encrypted: frames (visible page content — screenshots only)

The key exchange happens automatically when the viewer connects — no configuration needed. No input is accepted until E2EE is established.

Transport Security

• All WebSocket connections use WSS (TLS encrypted)
• Session tokens are short-lived (10 minute TTL by default)
• Each session is isolated — tokens grant access to one session only
• No credentials are stored or logged by the MCP server

Debug Log Safety

Debug logs (DEBUG=authloop:*) never contain:

• API keys or tokens
• Decrypted keystroke content
• Shared secrets or private keys
• Raw message payloads

Supported Input

Environment Variables

Debug namespaces

DEBUG=authloop:*           # everything
DEBUG=authloop:mcp         # MCP server + tool calls
DEBUG=authloop:session     # session lifecycle (create, connect, resolve)
DEBUG=authloop:stream      # WebSocket frames + input events
DEBUG=authloop:cdp         # CDP WebSocket commands/events
DEBUG=authloop:crypto      # E2EE key exchange
DEBUG=authloop:sdk*        # SDK HTTP client

Browser Compatibility

Works with any Chromium-based browser exposing CDP:

• Chrome, Brave, Edge, Chromium
• OpenClaw managed browser profiles
• Remote CDP (Browserbase, Browserless)
• Local or remote — HTTP endpoints auto-resolved via /json

Get an API Key

Sign up at authloop.ai — 25 free handoffs, no credit card required.

License

MIT