Computer use for AI agents: give an agent a computer your team shares
What computer use means for AI agents, the three ways to give one a computer (a cloud device, your own computer, one app), and who gets to start work on it.
Computer use is an AI agent working on a real computer: running commands, reading and changing files, and clicking and typing in apps, instead of only answering in chat. In Nebula an agent gets a computer one of three ways: a cloud device your workspace runs, your own computer with computer use turned on, or a single app on your computer through an app agent [S1][S2]. What decides which one to use is less about the agent and more about who else should be able to start work on that computer.
What “computer use” covers
A chat assistant hands you text and you do the work. An agent with a computer does the work itself. On a Nebula device that means:
- Commands and files. With computer use on, agents run commands from a working folder you choose, and their file tools reach either that folder only or the whole system, which takes a second click to switch on so access never widens by accident [S1].
- Your installed tools and private network. On your own computer, agents can use “your files, the tools you’ve installed, and anything only your computer can reach” [S1].
- One app’s window. An app agent reads what’s in one app’s window, such as Figma, Notes or Excel, and clicks, types and uses its menus. It never gets a shell, other apps or your files [S3].
- A terminal you can step into. When an agent on a cloud device hits a password or a sign-in code it can’t enter, the task shows a card and you type it into the device’s terminal yourself; the password doesn’t show on screen or in the task [S2].
Three ways to give an agent a computer
| Cloud device | Your own computer | App agent | |
|---|---|---|---|
| What the agent gets | A cloud machine for your workspace: Small (2 CPU, 8 GB RAM, 40 GB disk), Medium or Large [S2] | Commands and files from the folder you choose, with your installed tools [S1] | One app’s window and nothing else [S3] |
| Who can use it | You, people you name, the channels that use it, or the whole workspace if you share it [S2] | You, people you name and channels you add it to [S1] | Only you, in your Nebula DM [S1] |
| What the computer costs | Nebula Credits while it runs; a subscribed workspace starts with one small device that comes with the seat [S2] | Nothing [S2] | Nothing; it runs on your computer [S2][S3] |
| Runs when nobody is watching | Yes. Cloud devices never ask before scheduled work [S1] | Asks the owner first, unless they turn on Let scheduled work run on this computer [S1] | Only when you hand it work in your DM [S1] |
A new cloud device starts out Stopped and boots the first time an agent uses it, so that first task takes a little longer [S2]. A device can be upgraded one size at a time, and an upgrade is one-way [S2].
Who gets to start work on the computer
This is the part most “computer use” demos skip, because a demo has one person and one machine. A team has neither.
A channel’s agents work on the channel’s devices. Adding a device to a channel shares it with everyone in that channel: they can all run work on it there [S2]. When you create a channel on desktop, it starts with a new cloud device made for it; you can pick your own computer instead [S1].
Your own computer stays yours. It’s private to you unless you add it to a channel, name people, or set it to everyone in the workspace. If you add it to a channel the whole workspace can see, Nebula asks you to confirm first, because everyone who can use that channel will be able to start work on your computer [S1]. Workspace admins can’t grant access to your computer for you [S1].
Work nobody typed asks first on a personal computer. A scheduled task, a monitor, a message that arrived from Slack or email, or another agent handing work over pauses on a card that only the computer’s owner can allow [S1]. That is why work meant to run for the whole channel, around the clock, belongs on a cloud device.
How often an agent asks before it acts
Approval cards appear in the task itself. A command or file action outside the folder you allowed gets Approve and Deny; an action that needs your consent on that device can also get Always allow [S1]. For computer control you choose Ask each session, Ask every time or Always allow, per device, in advance [S1].
You also see when an agent looked. A screenshot or clipboard read through computer use raises a system notification, and a screenshot briefly outlines every display [S1].
Which one to pick
- The whole channel should be able to hand the agent work, including scheduled jobs: a cloud device. It’s always reachable and never stops to ask its owner [S1][S2].
- The work needs something only your machine has (a local build, a VPN, an installed tool): your own computer, added to a private channel or kept to yourself [S1].
- You want help inside one desktop app and nothing more: an app agent. Connecting the app is the go-ahead, and it can’t reach anything else [S3].
Coding agents such as Claude Code and Codex follow the same rules about where they run; running them in your team’s channels covers that in detail.
Common questions
What is computer use in AI agents? An agent operating a real computer, running commands, working with files and using apps, rather than only producing text for a person to act on.
Can an AI agent use my computer while I’m away? In Nebula, only if you allow it. Work you start yourself runs while computer use is on; with Allow using my computer the background helper keeps running when Nebula is closed. Work nobody typed asks you first unless you turn on Let scheduled work run on this computer [S1].
Do I need my own computer at all? No. A cloud device runs in the cloud for your workspace, and a subscribed workspace starts with one [S2].
Can one agent see every app on my screen? An app agent can’t: it only works in the one app you connected [S3]. Computer use with Screen Recording allowed can take screenshots, and each one raises a notification [S1].
Sources
- [S1] Nebula docs, Computer use — https://docs.nebula.gg/docs/computer-use (read 2026-10-31). Settings → Preferences → Computer use options (Allow using my computer, Allow only while Nebula is open, Don’t allow); working folder and Working folder only / Whole system access, “Switching to Whole system takes a second click, so access never widens by accident”; “agents can use your files, the tools you’ve installed, and anything only your computer can reach”; channel defaults to “New cloud device” with Use this computer as the alternative; confirmation before adding your computer to a workspace-visible channel; “Workspace admins can’t grant access to your computer for you”; app agent “Only you, in your Nebula DM”; approval cards and Ask each session / Ask every time / Always allow; work nobody started asks the owner, “Cloud devices never ask”; Let scheduled work run on this computer; screenshot and clipboard notifications.
- [S2] Nebula docs, Devices — https://docs.nebula.gg/docs/channels/devices (read 2026-10-31). Cloud device vs own computer table; “A subscribed workspace starts with one small cloud device, which comes with the seat”; cloud devices draw Nebula Credits while working, own computer costs nothing; sizes Small 2 CPU · 8 GB RAM · 40 GB disk, Medium 4 CPU · 16 GB · 250 GB, Large 8 CPU · 32 GB · 500 GB; one-way upgrades; new cloud devices start Stopped and boot on first use; adding a device to a channel shares it with the channel; terminal password and sign-in cards.
- [S3] Nebula docs, App agents — https://docs.nebula.gg/docs/computer-use/app-agents (read 2026-10-31). “An app agent uses one app on your computer for you, such as Figma, Notes or Excel”; reads the window, clicks, types, uses menus; “An app agent never gets a shell. It only gets the one app you connect.”