Platform

Computer

The Computer tab is where the agent does work that needs a working environment. Code, files, commands, builds, browsers, anything that needs to actually run instead of just be described. You watch it happen live and step in when you want.

Updated today

Overview

At a glance
Where
Computer tab in the right panel
Panes
Shell, files with editor, live preview
Activation
Opens automatically when the agent needs to run something
Control
You can take the keyboard at any moment
Persistence
Files and state stay across turns within a session
Restart
One click; resets the environment to a clean state

Most chat replies do not need the Computer. The Computer matters the moment a request stops being description and starts being execution. Build me this. Run this script. Try this fix and prove it works. Open this site and check the rendering. That is when the Computer comes alive.

What the Computer is

A working environment scoped to your session.

The Computer is a clean working environment the agent uses to perform tasks. It can read and write files, run commands, install packages, start servers, render previews, and do anything you would normally do in a terminal. It is isolated from your local machine so the agent cannot touch your laptop. It is scoped to your session so the work you do today is still here tomorrow.

Note
You do not need to set anything up. The Computer is ready the moment the agent opens it. If you want to bring in code from outside, paste a URL to a public repository or upload a folder.

Watching work happen

Three panes, all live.

PaneWhat it shows
ShellEvery command the agent runs and the streaming output of that command
FilesThe file tree, with badges for created, modified, and deleted files; clicking a file opens it in the editor
PreviewLive HTTP preview if the agent is running a server, image preview if it generated an image, video preview if it rendered one

All three update in real time. You do not have to refresh. Long-running work scrolls in the shell while the file tree updates and the preview reloads.

The shell view

The conversation between the agent and the working environment.

Every command the agent runs shows up here with its full output. Stdout and stderr stream live. You can scroll back through the session to see everything that happened.

Thing you can doHow
Read alongJust keep the tab open. The shell scrolls automatically when new output arrives
Search the shell historyCmd or Ctrl plus F in the pane to search within the visible output
Type a command yourselfClick into the input row at the bottom; the agent can see anything you type
Interrupt a running commandThe stop button at the right of the command row sends an interrupt
Copy outputSelect text and copy as you would in any terminal

The files view

Tree on the left, editor on the right.

The file tree shows the structure of the working directory. Files the agent created have a green dot. Modified files have a yellow dot. Deleted files cross out. Click any file to open it in the editor pane.

Thing you can doHow
Open a fileClick in the tree
Edit a file directlyClick into the editor and type; saves on blur
Create a new fileRight-click in the tree for the new file action
Rename a fileRight-click on the file in the tree
Drag a file in from your laptopDrop onto the tree to upload it into the working directory
Download a fileRight-click the file for the download action
Tip
Edits you make in the file tree are visible to the agent on the next turn. You can correct a small thing yourself and then ask the agent to keep going from there.

The live preview

Whatever the agent is rendering, shown live.

If the agent runs a dev server, the preview pane shows it. If the agent generates an image or a video, the preview pane shows that. If the agent builds a deck, the preview pane shows the deck. The preview reloads whenever the underlying artifact changes.

Preview kindRefresh behaviour
HTTP serverAuto-reload on file save (when the dev server supports it)
Static HTMLReload on the file tree change
ImageReplaces when the file changes
VideoReplaces when the file changes; click play to view
Console-onlyIf nothing renders, the preview pane stays empty and you watch the shell

Intervening

You can take over at any moment.

  1. 01
    Tell the agent to stop
    Hit the stop button on the running command, or send a chat message asking the agent to pause.
  2. 02
    Make your change
    Edit a file directly, run a command yourself, or upload something new.
  3. 03
    Hand back
    Send a chat message describing what you did and what you want the agent to do next.
Note
Direct user input in the shell or files is treated as authoritative. The agent will respect what you typed and continue from there.

What persists

The Computer is scoped to your session.

ThingStays until
Files in the working directoryYou delete them or you reset the Computer
Installed packagesReset of the Computer
Running processes (servers, watchers)You stop them, you reset, or the session ends
Shell historyAlways available within the session; cleared on reset
The preview URLStable for the lifetime of the session
Note
To carry work forward across sessions, ask the agent to save outputs to Brain or push them to a connected repository.

Use cases

Five things people actually do in the Computer.

Build a small app from scratch

Tell Sentinel what you want. The agent scaffolds the project, installs dependencies, writes the code, starts the dev server, and the preview comes up. You iterate by chatting: change the colour, add a button, hook it to this API. Every change you see live.

Fix a bug in your existing repo

Paste a public repository URL or upload your folder. The agent pulls it into the Computer, reads the code, locates the bug, writes the fix, runs the tests, iterates until green. You read the diff in the editor and copy what you need.

Generate a one-off script and run it

Ask for a script that does a thing (parses a log, processes a CSV, hits an API and writes a report). The agent writes it in the files pane, runs it in the shell, shows you the output. Download the output if you want it locally.

Build a slide deck or carousel and watch it render

For visual content the Computer runs the renderer and the preview pane shows the result. You scroll through slides, ask for changes, the renderer reruns, the preview refreshes.

Try an idea fast

The Computer is the lowest-cost way to prototype. Ask for a working version of an idea, watch it come together, decide if it has legs. If it does, push the code somewhere durable. If not, close the tab.