_desktopdocs
computer-use environment for agents

Give your agent a real computer.

A lightweight Linux desktop that runs entirely inside Docker — with an HTTP daemon and TypeScript SDK so any model can drive it: shell, GUI apps, browser, keyboard, mouse, screen.

Read the docsQuickstart
your-machine
$ git clone
  https://github.com/tanav29/desktop && cd desktop
$ docker compose up -d --build
# wait for healthy…
$ open http://localhost:6080/vnc.html

:6080 noVNC viewer  ·  :8095 HTTP daemon  ·  :8090 live stream

A real desktop

Debian 13 + XFCE + Chromium in one container, capped at 2 CPU / 3 GB. Xvfb on :99, noVNC in your browser.

The computer image →

One HTTP daemon

xdotool, ImageMagick and bash behind a clean API on port 8095. Shell quoting and process cleanup handled for you.

Daemon API →

TypeScript SDK

actions, cmd, create, kill, screenshot, live, frames — every call is one millisecond-level round-trip.

SDK reference →

Watch it work

noVNC viewer for humans, MJPEG stream + JPEG frame generator for agents and vision models.

live() and frames() →

Disposable by design

Boot with one command, reset with docker compose down. Only ./workspace persists — nothing to clean up.

Docker setup →

Model-agnostic

Plain HTTP end to end. Claude, GPT, Gemini, GLM or a bash script — if it can call an API, it can drive the computer.

Architecture →

Architecture

Three layers, each replaceable.

Every control surface is one HTTP round-trip to a process inside the container — milliseconds per call instead of ~150 ms+ spawning docker exec per action. And each computer is just a port.

  • Container

    Real GUI desktop: Xvfb → XFCE → x11vnc → noVNC. — Swap with any X11 desktop.

  • Daemon

    Python-stdlib HTTP API over xdotool / ImageMagick / bash. — Swap with anything that speaks HTTP.

  • SDK

    Thin TS client: actions, cmd, create, kill, screenshot, live, frames. — Swap with raw fetch() calls.

┌──────────────────── your machine ────────────────────┐
│                                                      │
│   web UI (Next.js) ────── eve agent                  │
│     (chat + live        (tools + loop)               │
│      desktop pane)                                   │
│                           │                          │
│                    computer-use-sdk                  │
│                           │ HTTP                     │
│                           ▼                          │
│   ┌──────────── Docker container ─────────────┐      │
│   │  daemon :8095   every call = 1 round-trip │      │
│   │    │                                      │      │
│   │  XFCE on Xvfb :99  ·  xdotool             │      │
│   │  x11vnc → websockify → noVNC :6080        │      │
│   │  Chromium · xterm · ImageMagick  │      │
│   └───────────────────────────────────────────┘      │
│                           │                          │
│              ./workspace persists on host            │
└──────────────────────────────────────────────────────┘
drive.tsfull API →
import { computer } from "computer-use-sdk";

// launch an app with a unique title
await computer.create("chromium https://example.com", {
  title: "web-1",
});

// a whole input sequence in ONE round-trip
await computer.actions([
  { do: "wait_for", window: "Chromium", timeoutMs: 45_000 },
  { do: "focus",    window: "Chromium" },
  { do: "key",      keys: "ctrl+l" },
  { do: "paste",    text: "https://news.ycombinator.com" },
  { do: "key",      keys: "Return" },
]);

await computer.screenshot("state.png"); // → workspace/state.png
const feed = await computer.live();     // watch it move
await computer.kill("web-1");           // dispose

Drive it from code

Batched actions. One round-trip.

Wait for a window, focus it, paste a URL, hit enter — in a single request. When a language model is the caller, one call per task beats one call per keystroke every time.

Run the smoke test

Boot your first computer.

One command, two minutes, zero host dependencies beyond Docker. Reset whenever you like.

Get startedTroubleshooting