Let your AI agent do real email work. Keep authority outside the prompt.
Connect Claude, Codex, ChatGPT, Cursor, or another compatible client to live contacts, messages, lifecycle flows, analytics, and product-enforced actions.
✓Human operators use OAuth; headless agents use scoped, expiring first-party keys.
Start with a read-only task
Operator prompt
Inspect the onboarding flow, explain where contacts stop, and show the evidence before proposing a change.
Connect
Remote MCP
Prompt
Inspect the live system
Boundary
Show before changing
[ 01 / 03 ]Agent safety
A useful agent needs context and a boundary.
Giving an agent a chat box is not the same as giving it a safe operating model. The connection must identify the principal, scope the tools, distinguish inspection from change, and preserve the evidence returned by every call.
Authenticate a principal
Authority comes from the human or agent credential, not from the chat client carrying it.
Scope the action
Read access, preparation, confirmation, and execution remain separate capabilities.
Retain the result
Tool responses expose identifiers and state so later work can be checked instead of guessed.
[ 02 / 03 ]Choose an agent
Choose the client. Keep the same product contract.
MCP clients
Setup for any client that speaks the protocol over HTTP.
Claude
Install the plugin and operate the live system with task-focused commands.
Cursor
Match product events in the repository to live templates and paused flows.
Codex
Compare application code with the runtime email system through one MCP block.
ChatGPT
Investigate delivery and prepare work through a compatible connector UI.
Hermes
Connect Hermes to Send & Retain through the shared tool contract.
OpenClaw
Give OpenClaw a controlled path to inspect and prepare email work.
Grok bot
Connect a Grok-powered bot to email operations through controlled tools.
eve
Use an internal eve agent to operate Send & Retain over its public MCP surface.
[ 03 / 03 ]Operating model
Move from evidence to a prepared change, then decide what goes live.
Connect
Use a compatible MCP client and authenticate a real principal.
Inspect and build
Let the agent gather context and build the change, paused.
Enable and observe
Turn the work on deliberately, and keep the evidence of what ran.
Shared guarantees
- OAuth 2.1 and scoped API keys
- No transport-specific credentials
- Confirmation-gated writes
- Auditable tool results
Questions people ask us
Connect one agent to the live email system.
Start with an evidence-backed inspection, then review one prepared change.