Letting an AI agent sign in

An AI agent that works in your browser sometimes reaches a sign-in page. With Vautir, it asks; you approve in Vautir’s extension; the extension fills the form. The agent never sees the password.

Set it up

1. Log the computer in with the command line, and start a session for the agent:

   vautir unlock --raw --minutes 480

   That prints the session’s key.

2. Add Vautir to your agent’s MCP servers, with that key:

   {
     "mcpServers": {
       "vautir": { "command": "vautir", "args": ["mcp"], "env": { "VAUTIR_SESSION": "…" } }
     }
   }

3. Keep Vautir’s extension unlocked in the browser the agent uses.

When the agent asks

Vautir’s button shows a count, and its popup asks: An AI agent asks to sign in to example.com, with the agent’s name, its reason, and your logins for that site.

• Fill it in fills the login into the tab open on that site.
• Refuse tells the agent no.

A request waits for two minutes.

What the agent gets

Which logins you have for a page, by title and username, and how its request ended: filled, refused, or approved with no form to fill. Never a password.

What to know

• Only the login’s own site. The extension fills only a page whose address the login matches. An agent that asks for one site’s login while on another gets nothing.
• The agent names itself, and gives its own reason. Read them as its claims.
• An agent that controls your whole screen could press the extension’s button itself.
• We see neither the agent nor the site. Each request is encrypted with your account’s keys.
• In an organization, a login filled from one of its vaults goes in its audit log.
• Only the browser extension answers agents.