Legato FAQ
Legato is an open-source (Apache-2.0) computer-use harness: it sends screenshots to a vision-language model, runs the actions it chooses on a desktop, and decides when the screen is ready for the next screenshot.
How is it different from Anthropic's computer-use demo? It runs any OpenAI-compatible model and waits for the screen instead of sleeping 2 s before every screenshot. On the same Claude model it was 1.48x faster on office tasks, at the same success rate (comparison).
Is it faster than OSWorld's Qwen agent? About the same speed, with more success: 71/96 vs 53/96 over two runs (benchmarks).
cua-agent and Bytebot are unmaintained. What now? Legato covers the agent loop for both: cua-agent, Bytebot.
Does it include a model, or need a GPU? No. Bring any model: Qwen through OpenRouter, Claude, or your own vLLM server.
Which desktops?
Full desktops: E2B Desktop, any sandbox through legato.attach(screenshot=..., execute=...), and OSWorld's Docker VMs
(sandboxes). For browser-only work, a Playwright-based tool such as Browser Use or Stagehand usually
fits better.
RL training?
legato rl collect runs graded rollouts with token logprobs and GRPO groups (rl.md).
How do I check it on my own tasks?
Compare result.report() from legato.Harness and legato.Harness.reference on them
(examples/compare_with_reference.py).
Is there a hosted version? Optional: legato.computer runs desktops with pause, resume and fork.