mailsocket — MCP server

An MCP server that exposes mailsocket to AI coding agents as native tools, so an agent can create a throwaway inbox and block on the OTP or magic link without writing any polling code. It wraps the Python SDK — no second HTTP client, no re-implemented polling.

What it does

Errors from the SDK (AuthError / NotFound / RateLimited / WaitTimeout) are turned into clean MCP tool errors — no stack traces, and the API key is never echoed back.

Reusing an inbox

create_inbox needs no since — a fresh inbox only ever has new messages. By default, wait_for_otp/wait_for_link omit since entirely, so the server matches only messages arriving AFTER the call (the request start time). But if you reuse an existing inbox (e.g. to request a second OTP/link), pass since explicitly: otherwise a call with no since will correctly ignore anything already sitting in the inbox, which is right for most reuse, but if you actually WANT to catch a code that already arrived before you called the tool, pass either the time right before you triggered the new email (ISO8601 or unix seconds), or the id of the last message you saw (msg_..., meaning "only messages after that one") — both tools return id/received_at on the matched message so you can chain since=<id> into the next call on that same inbox.

Trying the loop without real email

send_test_code does not go through the mail server, so the code is already in the inbox when the tool returns. Take a cutoff first: the id from get_latest (or since=0 if the inbox is empty), or the current time. Then call send_test_code(inbox_id), then wait_for_otp(inbox_id, since=<that cutoff>). Without since the wait only sees mail that arrives during the wait, so it would miss the test code.

Requirements

• Python 3.10+
• A mailsocket API key — from the dashboard or POST /api/v1/agents/register.

Install

No install needed with uv:

uvx mailsocket-mcp

Or install with pip:

pip install mailsocket-mcp
mailsocket-mcp

Either way, it reads MAILSOCKET_API_KEY (and optional MAILSOCKET_BASE_URL) from the environment. If the key is missing it exits with a clear message and status 2; the key is never logged.

Register with an MCP client

Claude Desktop / Cursor

Add this to your client's MCP config file — claude_desktop_config.json for Claude Desktop, ~/.cursor/mcp.json for Cursor:

{
  "mcpServers": {
    "mailsocket": {
      "command": "uvx",
      "args": ["mailsocket-mcp"],
      "env": {
        "MAILSOCKET_API_KEY": "ms_live_..."
      }
    }
  }
}

Claude Code CLI

claude mcp add mailsocket -e MAILSOCKET_API_KEY=ms_live_... -- uvx mailsocket-mcp

Any other MCP-compatible client works the same way: point it at the mailsocket-mcp executable (or uvx mailsocket-mcp) over stdio and pass the key through the environment.

Remote server (no install)

The same tools are hosted at https://mcp.mailsocket.app/mcp over Streamable HTTP. Send your key on every request as a header: Authorization: Bearer ms_live_.... Never put the key in the URL; a ?api_key= request is rejected.

Claude Code CLI

claude mcp add --transport http mailsocket https://mcp.mailsocket.app/mcp \
  --header "Authorization: Bearer ms_live_..."

Cursor (~/.cursor/mcp.json)

{
  "mcpServers": {
    "mailsocket": {
      "url": "https://mcp.mailsocket.app/mcp",
      "headers": { "Authorization": "Bearer ms_live_..." }
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "mailsocket": {
      "type": "http",
      "url": "https://mcp.mailsocket.app/mcp",
      "headers": { "Authorization": "Bearer ms_live_..." }
    }
  }
}

Any client that can set a custom HTTP header works the same way (OpenAI Agents SDK, the n8n MCP Client node, your own code).

Things to know:

• Claude.ai, Claude Desktop connectors and Smithery need OAuth, which is planned but not available yet. Until then, use the local stdio server above in those clients (Claude Desktop), or use a header-capable client.
• On the remote server, wait_for_otp / wait_for_link wait at most 55s per call, which keeps each response under proxy timeouts. On a timeout, call the tool again.
• The server is stateless: no session to keep, and every request is authenticated on its own. It has no key of its own and only acts with yours, under your normal API rate limits.
• Self-hosting: mailsocket-mcp-http runs the same server (see mailsocket_mcp/remote.py for its env vars).

Env vars

Develop

pip install -e ".[dev]"
pytest

Tests make no external or live-API requests. The SDK client is faked, and the remote-transport tests talk to a local fake upstream on 127.0.0.1.

Changelog

• 0.4.1 — requires mailsocket>=0.3.1: wait_for_otp / wait_for_link called without since no longer miss a message that lands between the SDK's internal long-poll calls (each follow-up call sends the cutoff the server reported in Mailsocket-Wait-Since). No tool signature changes.
• 0.4.0 — new send_test_code(inbox_id) tool (stdio and the hosted remote server, under the same per-key caps as the other non-wait tools): puts one sample OTP email into your own inbox without going through the mail server. Requires mailsocket>=0.3.0 (Client.send_test_code).
• 0.3.0 — BEHAVIOUR CHANGE (via mailsocket>=0.2.0): wait_for_otp / wait_for_link now omit since by default instead of forwarding 0 ("any message already in the inbox"). The old default could return a STALE OTP/link already sitting in a reused inbox; the new default matches only messages arriving after the call. Pass since=0 explicitly to keep the old behaviour. Hosted server (mcp.mailsocket.app, deployed from the repo, no new PyPI release): per-key concurrent waits raised 3 → 10 (= the Pro plan cap; the API still enforces each plan's own cap) and the shared wait pool 16 → 32, following the backend's raised wait capacity (global 64).
• 0.2.3 — wait_for_otp / wait_for_link accept an optional since (ISO8601, unix seconds or a message id) so a reused inbox returns the NEW code; results include the message id and received_at.
• 0.2.2 — depends on mailsocket>=0.1.3 (stricter _seg id validation in the underlying Python SDK).