Legato

Open-source computer-use harness for any OpenAI-compatible model: Qwen, DeepSeek, Claude, or your own vLLM server. It drives a desktop with a computer tool, waits for the screen instead of sleeping, and bounds every model call. Bring your own computer: your own screen (macOS, Windows, Linux), any sandbox you already run, or a cloud desktop.

Give your coding agent a computer

Claude Code, Codex, Cursor or any MCP client gets a desktop to look at and click: a container on your machine (needs only Docker), your own screen, or a cloud desktop (guide).

pip install "legatolabs[mcp]"
claude mcp add legato -- legato mcp --sandbox docker

_{A real run in real time: told that a revenue chart looks wrong, Claude Code opens the dashboard in a desktop container, sees the bars hanging from the top, changes one line and reloads to check the fix. Each image is a screenshot the MCP server returned.}

Or install the Claude Code plugin, which adds the server and a skill: /plugin install legato --marketplace legatolabs/legato. One click for other editors (these run the same server with uvx):

Build an agent in Python

pip install "legatolabs[e2b]"      # PyPI name: legatolabs. Import name: legato.

import os, legato
from e2b_desktop import Sandbox                       # needs E2B_API_KEY

agent = legato.Harness(model="qwen/qwen3.8-27b", base_url="https://openrouter.ai/api/v1",
                       api_key=os.environ["OPENROUTER_API_KEY"])
desk = Sandbox.create()
result = agent.run("Open the terminal and print the date", legato.attach(desk))
print(result.report())                                # where the time went
desk.kill()

legato.local() drives your own screen instead, legato.providers.docker(w, h) a desktop container on your machine, and legato.attach(shell=run) any Linux sandbox you can run a command in, such as a VM (sandboxes). With a model on your own GPU, nothing leaves your machine:

pip install "legatolabs[local]"
legato run "Open the terminal and print the date" --sandbox local --model Qwen/Qwen3.5-9B --base-url http://localhost:8000/v1

Results

Same model, same desktops, 24 office tasks per run. Method and raw data.

Model	Compared with	Time	Success
Claude Sonnet 5.5	Anthropic's reference loop	1.48x faster	48/48 vs 48/48
Qwen3.8-27B	OSWorld's Qwen agent	same	71/96 vs 53/96

How

• Waits until the screen stops changing instead of a fixed 2 s sleep.
• Up to 3 actions per turn, one screenshot after them.
• Drops old screenshots in chunks so the prompt cache keeps hitting.
• A wall-clock timeout and a token budget on every model call.

More

• Guides: a computer for your coding agent · test a desktop app · local model · office apps · why agents are slow
• Models · Sandboxes · MCP server · RL rollouts · FAQ
• Compare: computer-use MCP servers · Playwright MCP · Anthropic's demo · OSWorld's Qwen agent · Agent S3 · UI-TARS-desktop
• Migrate: from cua-agent · from Bytebot
• For coding agents: llms.txt · AGENTS.md

Apache-2.0. Hosted desktops and models: legato.computer.