Agent integration guide

Once Email Skill and MCP

Design safe, auditable agent workflows for email flows you own or are authorized to test. This page documents the boundary without claiming unreleased clients are available.

Integration contract published; product Skill and MCP Server not yet released
Review the API contractRead the testing guide
01

Safe automation scope

Create an isolated inbox, poll with a deadline, match a test marker, read the target message, assert the result, and delete the inbox. Every step needs a timeout, quota, and cleanup outcome.

02

Data that must stay out of model context

API keys, full addresses, verification codes, confirmation links, bodies, attachments, cookies, and session identifiers must not enter prompts, MCP logs, or ordinary analytics.

03

Skill responsibility

A Skill may orchestrate reviewed API operations, enforce redaction, and return a minimal test result. It must not replace authorization checks or enable account abuse, policy evasion, or third-party monitoring.

04

MCP Server responsibility

An MCP Server should hold credentials in a controlled runtime, expose narrow tools, isolate subjects, rate-limit access, audit operations, and support immediate revocation.

Recommended integration flow

  1. 1Store the API key in a secrets manager.
  2. 2Create one isolated inbox per test.
  3. 3Trigger mail only from an authorized system.
  4. 4Poll with backoff and a fixed deadline.
  5. 5Return only a redacted assertion result.
  6. 6Delete the inbox on success or failure.