Recover the project.
Read source-linked state instead of reconstructing a week of tabs. Every status names where it came from, how old it is, and what it cannot establish.
An evidence-backed command center for delegated technical work
Maat puts every project in one source-linked view, sends one finite job to the right model or machine, and counts work done only when a named gate returns evidence you can check.
The first job is recovery. Open one screen after a day away and answer what advanced, what is blocked, what died, and what deserves attention. The specimen below uses real registered project names and checked-in status labels. It is not a live observer.
Maat is being built around four capacities. The interface stays simple because the difficult machinery runs on your servers, under explicit authority, with a fixed end condition.
Read source-linked state instead of reconstructing a week of tabs. Every status names where it came from, how old it is, and what it cannot establish.
Tell Maat the outcome you want. It drafts a contract with one scope, one writer, one isolated worktree, one deadline, and an exit.
Route bulk work to your own machines. Convene independent frontier models when judgment matters. Keep cost in the server layer, not trapped inside an editor tab.
A worker cannot mark itself done. The named gate runs, known-wrong decoys must fail, and the result returns with a receipt that states exactly what passed.
Persistent autonomous writers create hidden authority. Maat gives a worker the smallest temporary power needed for one declared outcome, then takes it away.
Humans can override policy only when the override is explicit and attached to the exact work item. Conversation alone authorizes nothing. A panel vote proves nothing. The gate profile on the receipt is the claim.
A mixture of agents is useful when the agents can disagree. Each director answers alone, later rounds attack the strongest rival, and a fresh judge picks what survives. What follows is not a diagram. It is the actual panel run that designed this page.
Argued the page must start from the operator’s real problem: many projects, one screen, what changed since yesterday.
An event log cannot distinguish “nothing happened” from “nothing was watched.” The coverage line under the specimen above exists because this argument won.
Pushed the register and a self-sealing page as the hero. The judge cut that to a rim and kept its rigor: every honesty rule here is checkable against one data structure.
The seal-first hero. The inverted “do not request access” close. An absolute no-mocks rule. A five-verb feature taxonomy. What you are reading survived three written rounds.
The panel chose this page’s direction. It could not make the page true. The deploy, the tests, and the digest below did that.
A Maat mark never means “trust the AI.” It names the artifact, the gate profile, the scope, and the evidence. Change the profile and the claim changes with it.
A browser demonstration over a small canonical claims payload. It does not sign the HTML and it does not replay a product gate.
The product has a working research core and a public landing. The conversational command center is the next surface, and it stays labeled as vision until its action boundary passes hostile tests.
Maat is beginning with one demanding operator
If you direct many technical projects and need agents to work without letting fluent output become fake progress, ask to see the private build.
Open private Maat