Computer use for pi

Let an agent use the Windows desktop through PiCode's own desktop app: see a monitor or a window, read a window's controls, click, type, scroll, use the clipboard, open a program. The agent asks for one of twenty-three actions; PiCode checks that you switched that agent on, and the desktop app does the work on your desktop.

• Where: install pi-computer (Packages), then switch the agent on in Settings ▸ Computer (user menu, desktop app only).
• Not this: not the work browser. That reads and drives web pages inside PiCode's own browser. This page is the whole desktop.

Claude Code, Codex and the other agent CLIs get the same tool over MCP: PiCode tools for other agent CLIs.

pi-computer is an optional pi package — an extension, not part of PiCode core. It reaches the daemon over the same authenticated API every other call uses (the install token); it opens no port and adds no credential.

What it does, plainly

With the switch on, the agent acts with your own permissions, on your desktop, without a sandbox — the same way you would at the keyboard. Any window it can see, it can click; any program you can open, it can open. Off by default; off is one click away, and it holds from the next call. Text on the screen can try to instruct an agent; this version does not filter it. Keep the switch off for agents you are not watching.

What it can do

Action	Answer
screenshot, zoom, wait	an image the agent sees — one monitor, or one window when it names one
snapshot	a window's controls as lines: role, name, position in the last image, the text of a field
windows, focus	what is open, and bringing one window to the front
clicks, left_click_drag, mouse_move, scroll, type, key, hold_key	the action, then a fresh image
type with mode: "paste"	the text through the clipboard and one Ctrl+V — whole and fast, for long text or an app that garbles keystrokes; your clipboard text is put back afterwards (an image or files on it are not)
clipboard_read, clipboard_write	the clipboard's text
open	a program, a file or a URL, opened as Windows would

Every coordinate the agent sends is a pixel of the last image it received. The guidelines it carries say so, and say when to stop and ask you: before paying, sending a message, deleting files or typing a password.

Two things it needs

1. The desktop app, running and connected. The actions run inside the PiCode desktop shell on Windows. Without it the agent is told plainly: the desktop app is not connected.
2. A visible package and a switched-on principal. Install scope decides who sees the tool; the switch decides who may use it. A pi you started outside PiCode has no identity here and is always refused.

Managed agent, or a CLI in a terminal?

Where it runs	Identity	What it gets
An agent PiCode manages	its agent id	its own switch in Settings ▸ Computer
A CLI in a PiCode terminal	the terminal's id	its own switch, listed by the terminal's name
A pi you started yourself	none	refused, always

Where a click or a keystroke lands

Input goes to the window in front, and you move the front with every click of your own. So the agent may click, type or press keys only while the window in front is still the one it last looked at (its last screenshot) or chose (focus). If you clicked elsewhere in between, the action is refused and the agent is told to look again — nothing is typed into what you are using. The refusal is one of the rows on Settings ▸ Computer.

What you see

Every call, allowed or refused, is listed on Settings ▸ Computer and counted on the dashboard ("Desktop"). In the chat, each step shows the capture the agent got back as its last capture.