Skip to main content
Some agent work needs your recorded yes before it starts. Today that is the weekly walk of your own site: once a week an agent visits your domain and uses it the way a visitor would, to check that its pages, links and forms work for people and for AI assistants. It only browses. It does not sign in, buy anything, or submit a form that sends or changes something. What it checks, and what it files when something stops it, is on Signup check. Only a person signed in to the app. An API key or an MCP connection is refused, whatever its scopes, with a 403 that says:
consent is given by a person in the app
That holds on every door. An agent that tries is told where consent is given; it can never give it on your behalf. Each consent is one row in your workspace’s ledger. Rows are never edited, except to mark one withdrawn, so the history of who agreed to what stays readable. Under Settings › Autonomy (at /settings/autonomy), below the autonomy levels and the approval settings, is a row for your own domain, Agents on your domain. It says whether agents may visit it, and until when.
  • Read and agree opens the consent’s full text, word for word. Nothing is recorded until you click I agree; Not now records nothing.
  • Withdraw ends a live consent at once.
When a consent expires, the walk does not run. Renew it from the same row. The Scoreboard’s Site health view shows the walk as Needs you meanwhile, with a Give consent button that brings you to this page. That is your step; nobody at Attensira does it for you. Anyone in the workspace can withdraw a consent in the app. So can an API key or connection that may steer, because taking a permission away is always safe. No walk starts after that: consent is checked before each walk opens a browser. To end a walk that is already running, stop it. The walk is the only browser work agents do, and it is priced in credits: see Browser agents.

​
API

On MCP, list with type consent reads the ledger and act with revoke on attensira://{workspace}/consent/{id} withdraws one. grant is not an act action on MCP, and over REST POST /v1/actions with grant is 403 for every key and connection, whatever its scopes. There is no way to grant consent except a person in the app, and there never will be.

​
Passwords in a chat

Attensira never needs your password. Accounts are connected in the app under Customize › Connectors, where you sign in to the other service yourself. If you type a password into a chat, by any door (the app, email or Slack), it is removed before the message is stored. The message keeps the rest of what you wrote, and the password is replaced by a note that points to the connect flow. The agent never sees it, and it never appears in the chat’s title, its history or a run replay.