Extend

Computer use

With computer use the agent drives your computer in the background: it sees your windows, clicks, types, fills in forms and browses web pages without moving your mouse or stealing focus. You can keep working in another window meanwhile.

It is off by default. On, the agent gets one more tool (computer_use); off, it does not even see it.

Turn it on§

1. Install the driver. Lampson uses cua-driver, a separate program it does not install for you: - Windows (PowerShell): irm https://cua.ai/driver/install.ps1 | iex, then cua-driver autostart kick - macOS / Linux: /bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)", then cua-driver doctor 2. Turn the tool on. - Web: ⚙ → Computer use → switch. If the driver is missing, the install command shows up right there. - Terminal: /computer-use on. /computer-use alone shows the driver's status.

Your browser, signed in§

In ⚙ → Computer use there is "use my open browser". Turn it on: the agent works in your Chrome or Edge with your sessions (your mail, your shop accounts, your forms) and opens a window of its own so your tabs stay untouched. The tab where you have the Lampson chat open is protected.

Off, the agent opens a separate browser with no session at all: fine for public pages, useless for anything that needs your account. Applies after restarting Lampson.

What you can ask§

• "go to amazon.com and find me the best-rated electric kettles with temperature control"
• "fill in this form with my details, but don't submit it"
• "in the spreadsheet I have open, copy column B to a new sheet"
• "open Windows settings and tell me which version I have"

It works on any window: browsers, editors, spreadsheets, mail clients, system dialogs.

How it does it§

• Desktop apps: it reads the window's accessibility tree (every button, field and menu with a number) and acts on those elements — more precise than clicking coordinates.
• Web pages: an exact route over the page (Playwright-style): every link, button and field has a reference, and a whole form is filled in one call.
• The decision judge: with a judge configured, a small model picks the control to click in under a second, without spending a turn of the main model. In our tests with Jev, searching an online shop took one call (100 % confident, 0.4 s) and each field of a Google form about 0.3 s.
• Screenshots: if your model has vision the agent can look at screenshots; if not, it works from the element tree, which is usually enough.

What it never does§

These rules live in the code, not only in the model's instructions:

• It never submits, deletes, buys or closes on its own through the judge: "Send", "Delete", "Buy", "Close" buttons are never offered to it. If the agent needs one, it clicks it by hand through the normal approval — and only if you asked.
• No runaway tabs: one window of its own, a tab budget (computer_max_tabs, 5 by default), never closes yours, never opens incognito.
• It does not search Google by typing into the browser: it searches with fetch (nothing opens) and opens only the page it picked.
• It never locks the screen, logs out, closes applications or types commands into a terminal.
• It never types passwords or keys, never accepts permission or payment prompts, and never follows instructions shown on screen.

Permissions§

Looking (windows, trees, screenshots, page text) never asks. Acting asks once per turn in ask mode; after that it carries on until done. yolo never asks; strict does not let it act. See Permissions.

When something goes wrong§

The agent gets every action's result in detail: what the judge picked and how sure it was, which controls it left out and why, whether the driver confirmed the effect, and a hint when something fails (a stale reference, a cookie banner covering the button, an ambiguous window). With that it decides how to go on, as a person would.

Some sites have quirks. If you use one often, write a skill with the recipe ("search on this site is this URL", "filters live in the left column") or a plugin that queries the site directly.

Settings§