"Have your people call my people" finally has a machine equivalent. But handing your assistant the ability to contact real humans is exactly as dangerous as it sounds — unless the system is built so the assistant drafts, and you send.
This is the workflow RelayLink implements. Here's how it works end to end, and where the guardrails are.
Step 1 — Connect your assistant over MCP
RelayLink is an MCP server. Point your AI app at the /mcp endpoint with your personal API key and your assistant gains a small set of tools: check the inbox, read threads, draft a package, confirm a send. The key is yours, not the model's — whoever holds it owns everything sent under it.
There's no RelayLink app to install and no new inbox to check. Your assistant is the interface.
Step 2 — Work it out privately, then draft
You talk something through with your assistant the way you already do. When it's time to involve a real person, the assistant composes a briefing — the ask, the context, assumptions, options, decisions — from your session. The session itself never leaves your machine; only the briefing does, and only after the next step.
Step 3 — Review and confirm
Drafting and sending are deliberately separate. The draft sits server-side until you confirm it — a second, explicit step tied to your account. Two things happen at review time:
- You see exactly what the recipient will see. Nothing is added after your approval.
- Your note gets its provenance label. If you typed the words yourself, it ships as (verbatim, human-authored). If you approved the assistant's wording, it says that instead. Echoing a draft back is approval, not authorship — the labels don't lie on your behalf.
This is the part that keeps your assistant from "speaking as you." It can propose; only you can publish.
Step 4 — Delivery that asks nothing of the recipient
The recipient gets a normal email with the briefing rendered as readable text, plus a private link to a web view with a reply box. No account, no signup, no app.
- If they just reply — by email or through the link — their words come back into your thread, labeled as typed by them.
- If they have their own assistant connected, it can read the same briefing through MCP and help them think through a response privately before they answer. Symmetry is the point: both humans get a thinking partner; neither gets impersonated.
Every message they receive carries an unsubscribe link that blocks you in one click. Consent is enforced server-side, not promised in prose — more on that in the consent layer.
What to try first
A good first run is low-stakes and real: ask your assistant to brief a friend on a plan you've been mulling — a trip, a purchase, a project. Watch what it chooses to include, fix the assumptions it got wrong, and send. The reply arriving back inside your session — attributed, threaded, ready to reason about — is the moment the model clicks.
When you're ready, connect your assistant.