Power is earned in grades, never granted at once
Every action an agent can take sits at a grade. It starts read-only, earns approvals, and only whitelisted actions ever run on their own — some never do. MatePlug is in development; this page is the safety model it is being built to, not a feature list for a shipping product.
Looks things up, answers from your knowledge, and prepares drafts. Nothing is sent, nothing is changed, in any system.
Unlock: readiness check + replay test passed (see gates below).
Writes are queued as pending actions with the draft, the source data and the rule applied. A human approves each one — mostly from the phone — before it executes.
Unlock: 7+ days at L1, ≥80% of drafts accepted by your team.
Whitelisted actions run on their own, with sampling review afterwards. You can demote any action back to L2 at any time.
Unlock: 14+ days at L2, ≥98% approval success — and you switch it on.
The permanent exception: sending quotes and other money promises stays at L2 forever. No record, however clean, graduates a money promise to automatic.
The two gates
No write access before the homework checks out
Readiness check
Before an agent starts, MatePlug inspects what it will work with — data access, knowledge coverage, missing pieces — and reports back in plain language: what's ready, what's missing, what to fix first. No silent half-connections.
History replay
The agent re-runs your real past tickets and orders in a sandbox. Write access opens only at ≥85 overall, ≥90 on write actions, and exactly zero mistaken irreversible actions. Fail, and it stays read-only until it passes.
Hard rules
Lines the product does not cross
Writes go through the screen
Your data has dependencies only your own systems understand. So agents write through the system's own interface and validation — the same path your team uses. Databases are read-only side channels, never written to directly.
Everything is logged, everything reversible
Every read that matters and every action taken is written to a log you can inspect: what, when, from what data, approved by whom. Actions carry an undo path wherever the underlying system allows one.
Taking over ≠ approving
Agent Link may hold the mouse; that never grants it a decision. Screen control and action approval are two separate gates, and your own keyboard or mouse input takes control back instantly.
Recordings never train models
When you demonstrate a task, the recording is encrypted, kept for a short default period, and deletable by you. Learning means revising the SOP document — stated plainly, in writing, here: your recordings are not used to train models.
Credentials live in a vault
System logins you authorize are stored in a credential vault and used only for the connections you approved. They are never shown in chat, logs or exports, and a private connector generated for you never leaves your tenant.
When in doubt, it asks
An agent that can't verify a fact says so and hands over to a human with context attached. Guessing at an order status — or anything a customer will act on — is treated as a defect, not a style choice.
Built to be trusted with real work
Join the waitlist for early access — and bring your hardest safety question with you.
Join the waitlist