The agent browser

A real browser inside the desktop app that Perch drives for you, using the sessions you are already signed into.

Some work does not live in your folder. It lives behind a login: an invoice in a vendor portal, a filing on a court site, a dashboard your team pays for. The agent browser is how Perch reaches that work. It is a real browser built into the desktop app, running in a panel beside the conversation, and Perch drives it the same way you would.

It uses the session you already have

The important part is whose browser it is. Perch does not ask each service for API access or an OAuth grant. It works inside a browser session that you control and that you signed into yourself.

That matters for two reasons. Most of the tools people actually use do not offer an API worth having, and the ones that do often gate it behind a plan you are not on. And every OAuth grant you hand out is a standing permission you have to remember to revoke. A browser session is different: it is the same access you already have, doing the same things you would do by hand.

What Perch can do in it

Perch can navigate to a page, read what is on it, click, type, fill in forms, press keys, scroll, and dismiss the dialogs that get in the way. It can go back and forward through history and reload. It can also search the web and pull the results into whatever it is working on.

It reads a page two ways at once. It sees the structure, meaning the actual buttons, links, and fields, which is what lets it act precisely. And it sees the page as an image, which is what lets it cope when the structure is uncooperative and a control only makes sense visually. Most of the time the structure is enough. The screenshot is what keeps it working on the pages that would otherwise stop it.

One page at a time

The panel holds a single page. Perch navigates it rather than opening a pile of tabs, and links that ask for a new window open in the panel instead, so what you see is always what Perch sees.

Watching it work

The panel is not a black box. It opens on its own when there is something you should see, and the clearest example is a sign-in: if Perch lands on a login wall, the panel comes forward and tells you it needs you. You sign in, Perch keeps going from there. You never hand it a password, because you are the one typing it.

You can also just take the wheel. The panel has an address bar and back, forward, and reload, so you can drive it yourself at any point: go somewhere directly, get past a step Perch is stuck on, or check what it is looking at. When you are done, Perch picks up from wherever you left the page.

Where it draws the line

Perch will not click a button it cannot identify inside a dialog about sharing or permissions. Changing who can see something is not a step it will take on a guess, so it stops and asks you instead. This is deliberate: the browser is the one place where a confident wrong click has consequences outside your machine.

Where it works

The agent browser is part of the desktop app. The CLI does not have it, and neither does the web chat, because both of those lack the browser the panel is built on. If you ask the CLI to work in a browser, it will tell you the capability is not available there rather than pretend.

Getting started

There is nothing to turn on. Ask Perch to do something that needs the web and it will open the panel itself. Useful things to ask for:

Pull the last three invoices from a vendor portal and reconcile them against the folder.
Check whether this filing matches what the docket says.
Go to this page and get me the numbers out of the table.

If the site needs a login you have not done yet in the panel, expect Perch to stop and ask. That is the intended flow, not a failure.