Testing signup flows with an AI agent: give it a throwaway email + phone number

If you're building or testing an AI agent that needs to sign up for something — a SaaS trial, an app that requires phone verification, anything with an OTP step — you hit the same wall every time: someone has to babysit the inbox or the phone to relay the code back.

I built agent-identity-mcp, an MCP server that gives an agent both a disposable email address and a real phone number, so it can complete the whole signup loop itself — request the code, read it, submit it — without a human in the middle.

How it works

It's an MCP server with 5 tools:

• create_test_email — get a disposable inbox
• get_test_email_messages — read what landed in it
• list_services — see which apps/services have phone-verification support
• rent_test_phone_number — get a real number
• wait_for_test_sms — block until the OTP arrives, then return it

It bundles two things I already run in production: receivemail.dev for the email side, and sms-florin for the phone side — real UK carrier numbers, not VOIP, since a lot of signup flows now block VOIP ranges outright.

Setup

{
  "mcpServers": {
    "agent-identity": {
      "command": "npx",
      "args": ["-y", "agent-identity-mcp"],
      "env": { "SMS_FLORIN_API_KEY": "your-key-here" }
    }
  }
}

Get a key at https://flo-voice1.com/api-access. The email side needs no credential at all — mailbox creation is public and rate-limited.

Works with Claude Desktop, Cursor, Cline, or any MCP-compatible client.

Why I built it this way

The obvious alternative — mocking the email/SMS layer — doesn't actually test the real flow. Real providers reject disposable domains, real phone-verification APIs reject VOIP numbers, and both have delivery quirks that only show up with real infrastructure. This uses the real thing on both ends, so if the test passes, the actual signup flow works.

Repo: https://github.com/flovoice53-tech/agent-identity-mcp
npm: agent-identity-mcp

Happy to answer questions — curious what other pieces people wish existed for testing agent-driven flows end to end.