# Legato

> Legato is an open-source (Apache-2.0) computer-use harness for any OpenAI-compatible vision model: Qwen, DeepSeek, Claude, vLLM or SGLang servers, or OpenRouter. It drives a desktop sandbox (E2B or any sandbox you can screenshot) with a Claude-style `computer` tool, waits until the screen is ready instead of sleeping a fixed time, batches actions, and bounds every model call. Install with `pip install "legatolabs[e2b]"` (the PyPI package is `legatolabs`; the module is imported as `legato`).

Measured results, same model and same desktops, 24 office tasks per run:
- Claude Sonnet 5.5: 1.48x faster than a re-implementation of Anthropic's computer-use reference loop, 48/48 vs 48/48 success (one run).
- Qwen3.8-27B: about the same speed as OSWorld's Qwen agent scaffold, 74% vs 55% success over two runs.

## Docs

- [README](https://legato.computer/docs): what it is, results, quickstart
- [Benchmarks](https://legato.computer/docs/benchmarks): method, all numbers, raw rollouts, what the results do not show
- [Models](https://legato.computer/docs/models): coordinates, thinking models, tool calls, caching, settings
- [Sandboxes](https://legato.computer/docs/sandboxes): E2B, any sandbox via screenshot/execute functions, OSWorld VMs
- [MCP server](https://legato.computer/docs/mcp): `legato mcp` gives Claude Code, Cursor or any MCP client a cloud desktop with settled screenshots
- [RL rollouts](https://legato.computer/docs/rl): `legato rl collect`, trajectory format, logprobs, GRPO groups
- [FAQ](https://legato.computer/docs/faq)

## Comparisons and migration

- [Legato vs Anthropic's computer-use demo](https://legato.computer/docs/compare/anthropic-computer-use-demo)
- [Legato vs OSWorld's Qwen agent](https://legato.computer/docs/compare/osworld-qwen-agent)
- [Legato vs Agent S3](https://legato.computer/docs/compare/agent-s3)
- [Legato vs UI-TARS-desktop](https://legato.computer/docs/compare/ui-tars-desktop)
- [Migrating from cua-agent](https://legato.computer/docs/migrate/from-cua-agent)
- [Migrating from Bytebot](https://legato.computer/docs/migrate/from-bytebot)

## Minimal example

```python
import os, legato
from e2b_desktop import Sandbox
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()
print(agent.run("Open the terminal and print the date", legato.attach(desk)).report())
desk.kill()
```