hypervised desktop
for autonomous agents
Atlas gives your agent a hidden second desktop on your Mac. It opens apps, clicks and types there while you keep working on yours.
curl -fsSL https://atlas.insanearts.io/install.sh | shInside the harness.
Your agent’s desktop, in the tools you already use. Pick a harness and try a task.
“Find a free hour on Thursday afternoon and add a launch review to my calendar.”
Connect your harness in Atlas → Settings → Integrations.
Set up once.
Then it's your agent's Mac.
- 1
Install once.
One command, then the usual macOS prompts. Atlas sets up a hidden desktop on your Mac.
- 2
You keep working.
Ask your agent for something and get back to your work. Your screen, mouse and keyboard stay yours.
- 3
Watch it work.
Open Atlas to see your agent on its own desktop, clicking through your app and taking screenshots.
Bring what the task needs. Atlas can copy a build, sample files or settings into the helper, then clean up after the task.
Ask in plain language. Your agent opens apps, clicks controls by name, types and brings back screenshots.
Less context, same work.
Every screenshot and every page of instructions fills your agent’s context. Here is what the same three tasks took through Atlas and through each agent’s own Computer Use, on the same model.
ClaudeClaude Opus 5.5
6.4×less context with Atlas
Atlas36kComputer Use232kTokens by task Atlas Computer Use Write and save a TextEdit document 9.5k 37k Turn on three TextEdit settings 14k 25k Fill and submit a web form in Chrome 13k 170k - Screenshots a task
- 1.7vs 54.5
- Time a task
- 41 svs 203 s
- Tasks finished
- 9 of 9vs 8 of 9
Computer Use looks at the whole screen after every step, so a form with five fields took up to 153 screenshots.
CodexGPT-6.1 Sol, medium
1.4×less context with Atlas
Atlas20kComputer Use28kTokens by task Atlas Computer Use Write and save a TextEdit document 5.2k 9.3k Turn on three TextEdit settings 8.1k 11k Fill and submit a web form in Chrome 6.3k 7.9k - Screenshots a task
- 1.7vs 0
- Time a task
- 36 svs 30 s
- Tasks finished
- 9 of 9vs 9 of 9
Codex’s Computer Use reads apps as text, so it took no screenshots and finished a little sooner. With Atlas, Codex took about two a task. Most of the context gap is instructions: about 6k tokens, against 2k for Atlas.
Tokens of context each task added, median of three runs, on a MacBook Pro. Claude: Claude Code with Atlas, and Anthropic’s computer-use tool (computer_toolset_20260801, set up as its documentation describes) driving the same apps in the same helper. Codex: Codex with Atlas, and Codex’s own Computer Use on the desktop. Tokenizers differ, so each card has its own scale: compare bars within a card, not across them. A computer-use loop that takes fewer screenshots would cost less.
Quiet until there’s work.
The helper is a real macOS login on your Mac. Here is how it shares your Mac, and what it costs.
Idle means idle.
When your agent finishes, it quits every app it opened and leaves the helper quiet, so nothing sits in memory for no reason. Until the next task, Atlas does nothing.
One at a time.
Agents queue for the helper: one acts, the others are told Atlas is busy and wait their turn. Take Control puts you ahead of all of them.
You set the cost.
So it never drains your battery for nothing: watch live, every 5 seconds or paused, Stop holds every agent,
atlas close --allquits what they opened andatlas logoutsigns the helper out.
CursorIdle
- Atlas and the helper
- 1–4%of one core
- Extra power
- ≈ 0.3 W
- Battery an hour
- < 1%
Agent working nonstop
- Atlas and the helper
- ≈ 28%of one core
- Extra power
- 5–8 W
- Battery an hour
- 7–11%
top; power from the battery’s system power reading, over its 72 Wh for the battery column.Questions, answered.
A second macOS login on your Mac, signed in and running in the background. It is your agent’s desktop: its own windows, menu bar, Dock, settings and files. You stay in your own account and never see it unless you open Atlas.
Atlas helper Signed inNo. The helper is an ordinary macOS account on the same Mac, so apps run at full speed with no image to download or keep in sync. It shares your Mac’s hardware, network and speakers. It is not a security sandbox: give agents the same trust you would give them in a terminal.
No. Agents act only in the helper, through its own input service. There is no fallback to your screen. If the helper ever comes on screen, Atlas refuses to act until you switch back.
Computer Use works on your own screen: it brings windows up in front of you, you can’t keep working while it runs, and it takes over your screen, which hurts most on a small one. With Atlas the agent does all of that on its own hidden desktop, and your windows, keyboard and mouse stay yours.
Anything that needs a screen: open your build and check it in Dark mode, sign up on your local web app in Chrome, reproduce a bug someone describes, run an installer or a first launch, use Calendar, Notes or System Settings, and bring back screenshots of each step.
They take turns. Only one controller acts on the helper at a time; another agent is told Atlas is busy and waits. When you choose Take Control, your mouse and keyboard come first until you give it back.
Yes. Stop, in the menu bar or the viewer, cancels input underway and refuses every further agent action until you choose Resume. Agents can’t resume themselves.
Claude, Codex, Cursor, opencode and Pi. Claude works two ways: in the Claude app Atlas’s live view opens right in the chat, and in Claude Code /atlas-screen shows it. Codex and Cursor also show it in the chat; Cursor’s Agent CLI, opencode and Pi open Atlas’s own window. Connect them in Atlas’s Settings → Integrations, or with atlas agents install.
sample.csv copied inThe helper starts without them, and Atlas can keep your home folder private from it: the helper can’t list or open it, while your apps and other accounts keep their access. An agent copies in only what a task needs, such as a build or a sample file, and rolls it back afterwards. Browser profiles, keychains, app sign-ins, SSH keys and .env files are refused unless you turn copying sign-ins on, and nothing is copied back to your account.
Only the ones you sign the helper in to. In Settings → Accounts, Sign In to a Site opens it in the helper’s browser with agents paused and its screen hidden from them. Then choose how freely agents act there: Look Only, Ask Before Sending or Trusted. With Ask Before Sending they can browse and draft, and Atlas asks you before anything is sent: Just This Once, Allow for This Task, or Deny.
Not from Atlas. It has no telemetry or analytics. It goes online to install and for atlas update, which download from atlas.insanearts.io, and to show a listed site’s icon, which it reads from that site without your cookies or sign-in. Screenshots stay on your Mac; when your agent reads one, it goes to the agent’s model like any other file it reads.
Native Save and Open panels and Safari’s web views need a connected Screen Sharing session; Chrome and Firefox work without one. Drag and drop of files onto apps isn’t supported. The helper’s sound plays on your speakers.
Yes. Atlas is free and open source under the MIT license, and its code is on GitHub.
A Mac with macOS 14 or later and an administrator account for setup. Setup creates the helper account, or uses one you choose, and you sign in to it once to allow Atlas.
atlas update installs the latest release after checking its signature. atlas uninstall removes Atlas and keeps the helper account, which you can delete in System Settings.
Give your agents
a Mac of their own.
Install Atlas once. Your agent takes it from there.
curl -fsSL https://atlas.insanearts.io/install.sh | sh