Sign in to the agent page

Sign in to the agent page

Hand an agent a username and password to work through Personal Agent Proxy in a browser, with the same access settings, tasks and activity log as over MCP.

The agent page is Personal Agent Proxy for an agent that drives a browser rather than speaking MCP, such as Muse, Instinct or Dots. The agent signs in at your Personal Agent Proxy's address with /agent/login after it, such as https://app.personalagentproxy.com/agent/login, with a username and password of its own, and works as that agent login: it can do what the login is allowed and nothing more.

The page is one text field. The agent writes a script in it, the same the MCP server's script tool runs, chooses Send, and reads what the script came to under Answer. Every call a script makes is checked against the same access settings, and is in the activity log with the task it was made for. You can sign in with the same login yourself to try a script by hand, or find out why a call was refused.

Add the agent

On External agents, choose New agent, then the agent you use, and name it. Its login is made with a username and a password.

Hand the agent its login

Personal Agent Proxy shows a message to send the agent, with Copy message: the sign-in page, the username and the password. Instinct keeps passwords in a vault rather than in chat, so you send it a message asking for a link to a vault entry, enter the login on the page it sends, then tell it it is set up. The password is not shown again once you leave the page.

To hand it over again later, open the agent's page and choose Set up again, then New password. The old password stops working at once, and everything signed in with it is signed out.

Sign in

The agent, or you, enters the username and password on the sign-in page and chooses Sign in. A wrong username or password is refused, and so is the login of a revoked agent. A sign-in lasts a week.

The login's access is set as always, on its page under Permissions. A change counts from the agent's next request.

What the page shows

Under the text field, the page tells the agent what it would be told connecting over MCP: the connections the login can use, each with what it may do with it, the vault, and what a script can call: pap.startTask, pap.connections, pap.tools, pap.call, pap.finishTask and pap.feedback. A script is JavaScript or TypeScript run as the body of an async function, so one script can look in several accounts at once and answer with only what the agent needs. The answer is what the script logged, then what it returned, or the error it threw.

Tasks

Before anything else, a script starts a task with pap.startTask: a title in a few words and a goal, what you asked for in a sentence or two. The page keeps the task and sends it with the agent's later scripts, so their calls are made for it, until a script finishes it with pap.finishTask or starts another. A call made without a task is refused, saying to start one.

Runs are real

Everything a script does is done on your accounts, as it would be over MCP: sending an email sends it, and deleting a record deletes it at once. Each call is checked against the login's access as it is then, and is in the activity log as the agent's, with the task it was made for: allowed, denied by the access settings, or failed by the service, with what went wrong.

Sign out

Sign out, in the header, ends the sign-in and goes back to the sign-in page. Revoking the login on its page, or making it a new password, signs it out everywhere it is signed in.

A self-hosted Personal Agent Proxy has to be reachable from the internet over HTTPS for Muse, Instinct and Dots to sign in to it: they open the page from their own servers, not from your computer. Over plain HTTP, New agent offers them greyed out.