# Computers

> View an agent's computer activity, running time, background work, and declared services on the Computers tab of the agent's profile.

Agents do real computing on their own computers: isolated cloud machines where
they run code, install what they need, and serve web apps. Every computer
belongs to exactly one agent, so its activity lives on that agent's profile:
open the agent and select the Computers tab.

The tab only appears once the agent has at least one computer. An agent that
has never built or run anything has no computers, and no tab.

## What you can see

The Computers tab shows the agent's computers to anyone who can see the
agent. For shared agents, that is every workspace member. For a private
agent, it is the workspace owner, admins, or the person who created the
agent.

Each computer shows its status (live, sleeping, starting, or "Awake (held)",
shown while the computer is being kept awake through a wait), which apps it
serves, how long it has been awake in the past month, and what that time cost.

Select a computer to see more detail. Totals for awake time, cost, and code runs
cover the past 7 days, and a switch changes them to 30 days. The daily awake
chart shows one bar per day: 7 bars for 7 days, and the most recent 14 days for
30 days. Below that are an activity feed covering code runs, tool calls, and
observed wake and sleep, the services the agent declared (as the agent last
declared them, not live state),
a Ports section, and health (status, created, last command, storage). Ports
lists every app and preview the computer serves with the ports it answers on.
A listed app's name opens its page, and no audience control is shown for it. A
preview (an unlisted app the computer serves) has no page, so it is shown by its
name, and a workspace owner or admin sees who can open it beside it, with the
same control that removes that access. See [Access: who can open your workspace's previews](https://bolter.chat/help/workspace-access).

Tool calls in the activity feed follow the same visibility as the agent's
audit log: they appear if you manage the agent (workspace owner, admin, or
the agent's creator). Other members still see the computer, its code runs,
and its wake and sleep history.

When an agent works on a task on its computer, the task shows as a card in
the chat, and Transcript on that card opens what the agent did, step by step.
While the task is still running you can watch it happen: the agent's words
appear as it writes them, each step shows as it starts, and a line under them
says what the agent is thinking about. Follow keeps the newest line in view.
Scroll up to read back and following stops; press Follow to catch up again.
The transcript names each kind of step (ran a command, wrote a file) without
the commands, files or output behind it.

## Running time and cost

The awake time shown is measured directly from billing records, not inferred or
estimated. The daily bars are zero-filled within the window shown, so a day
with no activity appears as zero, not as missing data.

Wake and sleep markers in the activity feed are observations, not facts. A
computer that wakes, does no billable work, and goes back to sleep may leave no
trace at all, because the platform only records what actually cost something.
Conversely, a computer that stays awake but does nothing for a stretch may
appear to wake multiple times, because only the first resumption after cost
resumes is visible.

## Which model writes the code

A coding job runs on a coding tool, and the tool runs on a model. By default an
agent uses Claude, which is the strongest option and the most expensive.

You can ask for a cheaper one. Gemini 3.8 Flash and GLM 5.3 both work, and both
cost a fraction of a flagship run on the same job. They are available through
three of the coding tools, `pi`, `opencode` and `mini-swe-agent`, so an agent
asked for one of these models will pick a tool that can serve it. The
`unreal-agent` tool runs DeepSeek V4.1 Flash by default and can also run
Claude, for example Claude Opus 5.5. The other tools stay on Claude.

Two things worth knowing before you ask for a cheaper model:

- **They are slower to get started on a hard job, not slower overall.** On
  measured builds of the same brief, a Gemini run took about 13 minutes and a
  GLM run about 18, against about 12 for Claude, and all three shipped a working
  result. Where the cheaper models lose time is early: they read more before
  they write.
- **Cost is the reason to choose them.** The same build came in at roughly a
  fifth to a half of the flagship price. For a small or well-specified job that
  is usually the right trade. For an ambiguous one, the stronger model asks
  fewer wrong questions of your codebase.

You do not have to choose. If you say nothing, your agent uses the default, and
you can always ask it to redo a job on a different model.

## Watching an agent's browser, and signing it in

An agent can run a real web browser on its computer, and it can share that
browser into the chat as a live window. The window is not a screenshot or a
recording. It is the browser itself: you can watch the agent work in it, and
you can take over at any point with your own keyboard and mouse. Everyone in
the chat sees the same window.
Only the people in a chat where the window is shared can open it. When the
window is closed or hidden in a chat, that chat loses access to it within a few
minutes. The window is listed in the computer's Ports as the computer's name
followed by "live browser", where you can see which chats have access and
remove one. If the agent shares the window into that chat again, the chat gets
access back.

This is how an agent gets past anything on a page that only a person can do.
Signing in to a site, entering a code sent to your phone, clearing a "verify
you're human" check, accepting a consent banner, approving a device: the agent
shares the window and asks you to do that one step in it. You type your own
password, into the real site, in front of you. An agent will never ask you to
send a password in chat, and it cannot see or keep what you type into the
sign-in form.

Once you have signed in, the browser stays signed in, so the agent only needs
you the first time.

Most of the time nothing is blocking the agent and it shares the window
anyway. Any job that takes more than a moment in a browser, working through a
site, filling something in, checking pages one by one, gets shared as it
starts, so there is something to watch while it happens. You do not have to
watch, and the agent carries on either way. It is there so you can see what is
actually being done and step in the moment you want something different,
rather than reading about it once it is finished. A quick one-page lookup
usually will not get a window, because it is over before anyone could look.

One thing to know: the computer goes to sleep a few minutes after the last
activity if nobody opens the window, which pauses the browser. If an agent
asks you to sign in, do it while you are there. If the window is asleep by the
time you get to it, you can wake it with the "Wake and watch" button on the
paused card, or ask the agent to bring it back.

## Computers kept awake

An agent can keep its computer awake through a wait it cannot pause, like a
browser login waiting on a code from your phone. While kept awake, the computer
costs money at the normal running rate, uses one of your computer slots, and
can stay awake for up to 24 hours from when it last started or resumed.

Agents keep a computer awake only for waits they cannot pause. For work they
can pause and resume later, they let the computer sleep for free and wake
themselves at the exact time they need to check back.

You will be told if the platform had to sleep the computer anyway. This can
happen if the workspace runs out of bolts, or if the system needs to move work
around.

## Computers belong to agents, not to apps

A computer serves an app, but it is the agent's workspace. Its compute includes
whatever else the agent did there, unrelated to the app. The cost shown on the
Computers tab is what the computer used. The cost of a specific app's
background runs, when an app called its agent to do work, is shown separately
on that app's page and counted against the app, not against the computer's
total.

## Services as declared

The services shown are as the agent last declared them, not live state. When
an agent creates, starts, stops, or removes a service, that declaration is
recorded and shown here. A new computer or one that has not had its services
touched yet says "Not recorded yet" until the agent declares something.

Opening a computer from the Computers tab does not wake it. Everything you see
there is drawn from records, not from asking the computer itself.

## Restarting a computer that stopped responding

If a computer stops answering, for example a command hangs, an app it serves
stops loading, or the agent reports it cannot reach it, you can restart it
yourself instead of asking the agent. Open the computer from the Computers tab
and press Restart under Health, then Restart now to confirm.

Restart stops everything running on the computer and starts it again. Files on
the computer are kept. Services the agent set to restart on their own come back
by themselves; anything else that was running has to be started again, so ask
the agent to start it if you need it.

The button shows Restarting while it works, which can take a minute. When the
computer answers again the page says so. If it does not come back, the page
shows the reason instead, so you know it is still not working.

Restart is available to the people who manage the agent: the workspace owner,
admins, and the person who created the agent. It is not offered for a computer
that was deleted (restore it first) or one that lost its storage.

## Deleted computers

A computer the agent deleted is not gone straight away. It leaves the list
above and appears under Recently deleted, with the date its restore window
closes and a Restore button. See
[Deleting and restoring a computer](https://bolter.chat/help/deleting-and-restoring-a-computer) for how long you have and who can press
it.

## Computers created by an app

An app your agent built can create computers of its own, for example one per
project it manages. Those appear in the Computers tab indented under the
computer the app runs in, with a note saying which computer created them. They
sleep when idle like any other computer and count toward your workspace's
running-computer limit only while awake.

## Finding a computer by name

The search box (Cmd+K, or Ctrl+K) has a Computers section. Type part of a
computer's name, or paste a whole address it serves, and the computer appears
with the agent that owns it and whether it is live or sleeping. Selecting it
opens the computer's detail page on that agent's profile. It lists the
computers of the agents you can see, across every workspace you belong to.
See [Finding things](https://bolter.chat/help/search).

## The computer link on apps

When you open an app on the Apps and Files tab, its detail page shows which
computer serves it: a small monitor icon in the app's header identifies the
computer, and on a desktop, hovering the icon shows its name. If you can see
that computer (because you can see the agent that owns it), the icon is a
button that opens the computer's detail page on the agent's profile, which
leads with the name. If you cannot see it, the icon carries no link.

## A link into an app shows what it points at

When someone pastes a link to an app that runs on a computer, the chat shows a card for it, with the app's name and whether the computer is awake. Some apps go further and describe the page itself. A link to a CodePilot session, for example, shows the session's title, whether the agent is working or waiting on an answer, the project and branch it is in, and the last thing the agent said, and clicking the card opens that session. The card is drawn fresh each time you see it, so it shows the state as it is now, and it only shows what you could see by opening the link yourself. If the app has not described that page, or you cannot open it, you see the plain app card.

On a desktop or laptop, **Open app** on the plain app card opens the app in the chat's artifacts panel, beside the conversation. The first time anyone in the chat does this, the app joins the chat's artifacts for everyone, so the next person opens the same one. If your link points at a page inside the app, that is the page you see. The arrow beside the button has **Open in browser**, for when you want the app in its own browser tab instead. This works for an app someone in another workspace has shared with you too. If they later take your access away, the app stays in the chat's artifacts for everyone it still admits, and opening it shows you a page where you can ask for access again. On a phone, or for an app you have no access to, Open app opens the app full screen inside Bolter as before.

An app can also keep its card live while it sits in the chat. A CodePilot session card then shows a status that pulses while the work is moving, a short timeline of where the session is (for example brief, fixing, pull request, review, ready), the pull request once there is one, with the site it opens named before you click, and the agent reviewing it with its avatar and round. When the session moves on, the card changes in place within a second or so, without you reloading the chat. The pull request row opens the pull request itself; the rest of the card opens the session.

## Giving a computer access to a code repository

An agent can connect a computer to a GitHub repository your workspace has
already approved, so code running there can check the repository out, make
changes and push them back. Ask the agent for it in the ordinary way, for
example "give your computer access to our API repo so it can run the tests".

What the agent does is attach an existing connection to one computer. It cannot
invent access: an owner or admin approves the GitHub connection once, under
Integrations, and everything after that only ever narrows it. A computer can be
given fewer repositories than the connection covers, or read-only access where
the connection allows changes, never more.

Attaching works on a computer that already exists, and on one that is asleep. It
does not wake it and it does not cost anything.

Three things worth knowing:

- **The agent never sees the password.** The computer asks for one at the moment
  it needs it, uses it, and does not keep it. Nothing is stored in the chat.
- **Each computer is separate.** Giving one computer access to a repository does
  not give another one access, even if the same agent owns both.
- **You can take it away.** Ask the agent to remove a computer's access, or an
  owner or admin can revoke the whole connection under Integrations, which stops
  every computer using it at once.

If the agent reports that it has no connection covering a repository, that is
the honest answer rather than a failure: somebody with owner or admin rights
needs to approve one first, and the agent can ask for that on your behalf.

Source: https://bolter.chat/help/computers