Webfold
Menu

Product · Computer use

A sandboxed computer. Your agent drives it. A person can take over.

Give an agent an isolated desktop. It controls it through the accessibility tree. You watch it live in a browser. You can take over whenever you need to. Every change of hands is recorded.

Get startedJoin the waitlist

See it work

Try it from wherever you work.

1cargo build --release -p computer-gateway
2computer-gateway --print-default-config > gateway.json
3computer-gateway --config gateway.json

What you get

What Computer use does.

Structural, not pixel-guessing

The agent reads the desktop through the accessibility bus and acts through accessibility interfaces first, so it works on Wayland without screenshots.

Watch, take over, hand back

A person can watch live in any VNC-in-the-browser client, take input, and give it back. Exactly one party can type at any moment.

Evidence you can trust

Every change of hands and every action is recorded in a tamper-evident, hash-chained audit log.

Three isolation tiers, and Android

Process, container and micro-VM backends, plus an Android backend with a UI tree.

How it works

Three steps.

  1. 1

    Create

    The gateway starts a hardened desktop session.

  2. 2

    Drive

    The agent connects over MCP, a JSON line adapter, or a token-gated local socket.

  3. 3

    Supervise

    Watch the stream, take over if needed, and read the audit log.

Limits up front.

Know the limits

  • This is a first working slice. The sandbox docs label each control as enforced and tested, enforced but untested, or documented only. Read them before relying on a control.
  • The container and Nix packaging were written for a standalone workspace and are not yet verified in the merged tree.
  • You can run the gateway on your own Linux host, or create computers through the hosted API's /v1/computers endpoints, which forward to a gateway for your account (early access). The live desktop stream connects to the gateway directly.

Keep reading.

Computer use guide →

Browser →

Agents →

FAQ

Questions.

Does it use screenshots?

No. Perception comes from the accessibility tree. Actions go through accessibility interfaces first and fall back to a virtual input device.

Which faces are there?

An MCP server, a JSON-line adapter for computer-use clients, and a token-gated local socket face with rate and time budgets.

Use Computer use with Webfold.

Run the command-line tool today. Join the waitlist for the hosted API.

Join the waitlistRead the quickstart