What it does
A self-hosted gateway for the A2H (Agent-to-Human) protocol, shipped as a single Go binary with SQLite. An agent or script sends a request for approval, a form or information, and a person answers by email, Telegram, push notification or a paired phone app. The agent receives a signed answer that records who decided. The gateway covers the A2H 1.0 agent-facing API, including webhooks with HMAC signatures and polling as a fallback, JSON Schemas for every message, and a conformance suite for testing other A2H gateways. It is an independent implementation, not affiliated with Twilio.
It suits workflows where an AI agent should check with a person before acting, such as deleting files or spending money. When high assurance is requested, the approval must be confirmed with a passkey (WebAuthn). Replay protection relies on message IDs, a timestamp window, single-use nonces and single-use links. Agents in Claude Code, Claude Desktop, Cursor and other MCP clients can use the built-in MCP server (`a2h mcp`), which provides tools to request approval, ask for a short form, send a notification, check a pending request and cancel one. The recipient is fixed in configuration, so the model cannot choose whom to contact.
To try it, download a binary, use the Docker image or install with Go, then run the gateway and open the link it logs to approve a test request. For production, place it behind HTTPS and set the public base URL. Channels are configured through environment variables for SMTP, a Telegram bot token, or ntfy push, and agents authenticate with API keys sent in a request header.
Limitations and requirements
Production use requires HTTPS, because people open the gateway's links and passkeys only work on a secure origin. The public base URL variable is required. Telegram needs a bot created with BotFather and a first message sent to the bot so it can write back; action buttons appear only when the base URL is a public host name. Passkeys can only be registered by the operator on the server. Development mode disables authentication and must never be exposed. A tool call waits up to 10 minutes before the agent must follow up with a status check.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
A2H: ask a human
Ask a human before acting: approvals by email, Telegram or push, signed answers, passkeys.
v0.3.0As shown on registry.modelcontextprotocol.io when we last read it