---
title: Computer chats
description: Chat with an agent that works in one of your Venom Computers, watch its live screen beside the conversation, and answer the browser's asks in the chat.
section: agents
order: 7
audience: [user]
sources: [services/nodejs-cloud/lib/agents/index.ts, services/nodejs-cloud/lib/computers/agent-pure.ts, services/nodejs-cloud/components/agents/ChatView.tsx, services/nodejs-cloud/components/agents/ComputerPane.tsx, services/nodejs-cloud/components/agents/screen-pane.ts, services/nodejs-cloud/components/computers/LiveScreen.tsx, services/nodejs-cloud/components/computers/control.ts]
features: []
---

A computer chat is a chat with an agent that works in one of your organisation's [computers](/docs/computers.md): a
persistent browser on one of your daemons, with the logins it already has. Use it for work in a website rather than in a
repository: filling in a form, checking an order, collecting numbers from a dashboard you are signed in to.

## Start one {#start}

On a new chat, open **Where** and pick the computer under **Computers**, then type what you want. No project folder is
needed.

Don't have a computer yet? Pick **My computer (new)**. It is a browser computer that belongs to you: the cloud makes it
the first time you pick it, on the daemon with the most room, and later chats reuse it, so the sites you log in to stay
logged in. It is listed in Venom Computers under your name ("ana's computer"), and once it exists it is first under
**Computers**. The computer starts with the chat, so the first answer can take a little longer. When no daemon offers a
project folder, a new chat picks My computer by default. Through the API or `agents_chat_create`, send
`{"kind": "computer"}` with no `seat_id` for the same thing ([Chats API](/docs/agents/api.md)).

You can start a computer chat when:

- your organisation has both Venom Agents and [Venom Computers](/docs/computers/overview.md#the-add-on) (both are included in
  Team and up);
- your role may use Venom Computers (owner, admin or developer by default), and you may use this computer
  ([Who can do what](/docs/computers/overview.md#who-can-do-what));
- the computer is placed on a daemon and is not archived. Otherwise the chat is refused with `seat_unplaced` or
  `seat_archived`. For My computer, a daemon with Computers turned on must have room for one more (`no_capacity`
  otherwise).

The agent runs in the computer's own session on its daemon, under your [member key](/docs/agents.md#your-usage-your-limits),
with the model and mode you picked.

## Watch it work {#screen}

The computer's live screen sits beside the conversation, so you see each page the agent opens and each click. Watching
needs nothing: the screen goes live as soon as the computer runs, including when the agent starts it with its first step,
and reconnects by itself after a drop, after the computer wakes and when you come back to the tab. On a phone, switch
between **Chat** and **Screen** at the top; the Screen tab shows a dot when the agent needs you.

Two buttons in the screen's header change its size:

- **Expand** (four corners) shows the screen as large as the window allows, in full screen where the browser supports it,
  with Take control, Release and the "needs you" banner still at the top. Use it to type into a form or to scan a QR code
  that is too small beside the chat. Press `Esc` or the same button to come back. It works on a phone too.
- **Minimise** (desktop) folds the screen to a slim bar at the right edge so the conversation gets the width; the bar shows
  a warning sign when the agent needs you, and one click brings the screen back. The browser remembers whether you left it
  minimised.

Above the screen, with the same rules as the computer's own page
([Watch and take control](/docs/computers/watch-and-control.md)):

- **Take control** gives you the computer, for example to type a 2FA code: your clicks, scrolling and keys go to it,
  and the agent waits. Only one person controls a computer at a time. You keep control while the chat is open; it is
  handed back when you press **Release** (or `Ctrl+Alt+R` in the screen), close the chat or leave the page.
- While you control a browser computer, an **address box** sits above the screen: type a site or a full address and
  press **Go** to open it.
- When the agent asks for a person (a login, a code, a CAPTCHA), a banner says so. Take control, do the step, then
  press **I'm done**, or **Skip** to let the agent carry on without it.
- An asleep computer shows **Start**.
- **Open in Computers** opens the computer's own page, with its activity and settings.

## It asks before it acts {#asks}

The computer's browser asks a person before the first action on a site, before it touches a credential or payment field,
before a submit that sends something, and on foreign frames, uploads and downloads. In a computer chat each ask appears
in the stream as buttons, like a code chat's approvals. Sites a person already allowed on this computer (**Allow on
example.com** in an [agent run](/docs/computers/agents-on-computers.md#it-asks-before-it-acts)) are allowed in the chat too.

With the **Auto** mode the agent opens and uses sites in the computer's browser without asking: the browser is the
computer's own, in its own container on the daemon. It still asks before a credential or payment field, before a
submit, and on a frame it can't inspect, and your organisation's policy and deny rules still apply. Pick Auto while an
approval to open a site is waiting and that approval goes through on its own.

## In the sidebar {#sidebar}

Computer chats are grouped under their computer in the sidebar, beside your projects. Archiving one stops the agent and
closes its session; the computer itself keeps running, with its logins.

## Related {#related}

- [Agents on computers](/docs/computers/agents-on-computers.md): give a Chat agent or a Tasks agent a computer instead of
  chatting with one yourself.
- [The Agents app](/docs/agents/app.md): everything else about the chat.

## The model

A computer chat's model runs on the computer's daemon or on your organisation's cloud keys, as **Model runs on** in the
new-chat box decides, like a code chat's ([Where a daemon chat's model runs](/docs/agents/model-source.md)). With cloud keys the
screenshots and pages the agent reads go through Venom Cloud to the model's provider.
