Computer-use agent

An AI agent that controls your computer

Not a chatbot that describes what to do. An agent with a shell on your machine, your real browser with your logins, and a list of tasks it works through. You choose which tools it may touch, which it must ask about, and which are off limits.

~90 KB
bridge binary (Linux build)
5
model vendors on one plan
3
permission levels per tool
Open
source edge and CLI
Definition

What "controls your computer" actually means

Most AI products run in a vendor's cloud. They can browse in a sandbox and produce a document, but the moment a task needs your files, your installed tools, or an account you are already logged into, they stop and hand it back to you. A computer-use agent removes that hand-off: it has a shell on the machine where your work lives and a browser where you are already signed in.

TODO for AI does this in three places at once. A native bridge on your PC, Mac or Linux box. A Chrome extension in your real browser profile. And a cloud VM per account with its own cloud browser, for work that should keep running when your laptop is closed. A single task can move between them.

Setup

How it gets onto your machine

Three optional pieces. Install none of them and it still runs in the cloud VM. Install all three and it can reach everything you can.

01
Bridge

A native binary, about 90 KB on Linux and 150 KB on Windows. Noise-encrypted WebSocket, PTY relay. Gives the agent a shell on that machine. Runs on Linux, macOS and Windows.

02
Chrome extension

Drives the browser you already use, in the profile that is already logged in to your CRM, ad account, bank portal or admin panel. No API keys, no OAuth apps to build.

03
Edge and CLI

Open-source runtime that ships about 90 CLI tools as integrations and wraps any binary you have. Also what the cloud VM runs.

Use cases

What people hand it

The list is the interface. Drop a task, the agent picks the machine and tools it needs, does the work, and shows the result with the receipts.

01
Repo work on your box

Fix the bug, run the tests, open the PR, from the checkout on your own disk with your own toolchain.

02
Logged-in browser chores

Update the CRM, pull the invoice PDF from the vendor portal, post the ad, reply on the support desk. Anything that needs a session you already have.

03
Files and documents

Rename and sort a download folder, build an XLSX with formulas, turn notes into a DOCX, merge PDFs.

04
Scheduled and unattended

The weekly numbers, the daily triage, the backup check. Runs on the cloud VM whether your PC is on or not.

Safety

Staying in control of an agent with a shell

Every tool an agent can use can be set to allow, ask first, or block, per agent. Secrets live in an encrypted vault; connected tools keep their own sessions so they stay signed in between tasks. Any run can be stopped mid-task. Cloud sandboxes are Firecracker microVMs, one per account, hosted in the EU.

The honest limit: an agent with a shell can do damage if you leave everything on allow and walk away. Set ask-first on anything that writes, and widen from there.

Decision guide

Which tool for which need

Runs on my own PC with my installed toolsTODO for AI, OpenClaw, Hermes Agent
Acts in my real logged-in Chrome profileTODO for AI, Claude Cowork (Claude in Chrome), Hermes Agent
Keeps running when my laptop is closed, with no server to runTODO for AI, Claude Cowork, Manus
Any model vendor on one planTODO for AI
Free and fully self-hosted, bring your own keysOpenClaw, Hermes Agent
Vendor-managed sandbox, no local accessClaude Cowork, ChatGPT agent, Manus
Our take
Cowork and ChatGPT agent run on their computer. OpenClaw and Hermes run on yours if you operate them. TODO for AI runs on yours and on a managed cloud VM, in your real browser, with any model, as a product you do not have to operate.
FAQ

Common questions

Is an AI agent that controls my computer safe?
As safe as the permissions you give it. TODO for AI lets you set every tool to allow, ask-first or block, keeps secrets in an encrypted vault, isolates cloud runs in Firecracker microVMs, and lets you stop any run. The edge runtime and CLI are open source so you can read what they do.
Does it work on Windows and Linux?
Yes. The bridge and desktop app ship for Linux, macOS and Windows. Claude Cowork is desktop on macOS and Windows only; TODO for AI is the option for Linux.
Do I need API keys?
No. One subscription covers Claude, GPT, Gemini, Grok and Ollama. For tools, the Chrome extension uses the sessions you already have, so there is nothing to wire up.
How is this different from Claude computer use or ChatGPT agent?
Those drive a browser in the vendor's sandbox. TODO for AI drives your actual machine and your actual Chrome profile, plus a cloud VM you own, and does it with whichever model you pick.
How is this different from OpenClaw or Hermes Agent?
Same idea, different packaging. They are free, MIT and self-hosted: you run the gateway, hold the keys and keep the box on. TODO for AI is hosted, adds a persistent task list, a managed cloud VM and browser, a vault, team roles and per-tool approvals.