Legato vs Anthropic's computer-use demo: which loop should you use?

Short answer. Use Anthropic's computer-use demo if you want Claude with bash and file-editing tools, Anthropic's own computer tool versions, or Bedrock and Vertex, in a Docker desktop you can watch. Use Legato for the same screenshot-and-click loop with less fixed waiting, on a sandbox you already run, or with a non-Claude model. On 24 office tasks with Claude Sonnet 5.5, Legato took 0.678 of the reference loop's agent time (95% CI 0.604 to 0.757; 1.48x faster), with 48/48 successes in both arms, in one run against our re-implementation of the demo's loop.

At a glance

	Anthropic computer-use demo	Legato (`legato.Harness`)
Models	Claude only: Anthropic API, Bedrock or Vertex [1]	Any OpenAI-compatible chat API that accepts images; Claude measured through OpenRouter [8]
Desktop	Ubuntu in Docker; the loop runs inside the container it controls, one session at a time [2]	E2B Desktop, or any sandbox given a screenshot function and an execute function [9]
Wait after action	2.0 s sleep before every screenshot [4]	Waits until the screen stops changing, capped at 2 s per action (longer for `wait`) [10]
Screenshots kept	All on the Anthropic API (caching turns truncation off); last 3 by default on Bedrock and Vertex [5]	Last 8; older ones dropped 4 at a time so the cached prefix holds [8]
Actions per turn	Every tool call in the reply, with a screenshot after each [6]	Up to 3, with one screenshot after them [8]
Call bounds	4,096 output tokens; 4 SDK retries on the Anthropic API; no explicit timeout; no step limit [6]	4,096 output tokens with a cut-off note; 60 s timeout; 1 retry; 30 turns by default [8][11]
License	MIT [7]	Apache-2.0
Status	Active (commit 4a4224b, 6 November 2026)	Version 0.1.0

What we measured

Legato Office Bench: 24 tasks that move data from LibreOffice Calc into Impress, 2 attempts each, 25-step cap, real Ubuntu desktops, both arms at once on one host, Claude Sonnet 5.5 through OpenRouter, 1366x768 screenshots and prompt caching in both arms [12].

Per task	Legato	Reference loop
Agent time (mean)	65.7 s	96.9 s
Model time	42.4 s	44.1 s
Waiting for the screen	14.0 s	44.5 s
Model calls	14.0	10.6
Input tokens per call	8,603	18,667
Success	48/48	48/48
OpenRouter cost	$0.126	$0.137

Almost all of the saving is fixed waiting. Legato makes more calls, but each is smaller and shorter (3.0 s vs 4.2 s).

Limits of this result:

The reference arm is legato.Harness.reference, a re-implementation. It keeps a text history instead of tool_result blocks, is given the first screenshot (this favours the reference), offers only the computer tool, and caches through OpenRouter's top-level cache_control instead of per-turn breakpoints. We have not run the unmodified demo in this benchmark.
One run, office tasks only. The pre-set "substantially faster" bar (ratio 0.67 or less) was missed by 0.008.
This run used Legato's earlier 1,024-token output budget. The current default (4,096 tokens and a cut-off note) has not been re-measured on Claude.

When to use Anthropic's demo instead

You need a shell or file editing. The demo gives Claude bash and str_replace_based_edit_tool [3]; Legato offers only the computer tool.
You want Anthropic's computer tool itself. The demo sends Anthropic-defined tool versions, up to computer_toolset_20260801, and Anthropic's thinking modes. Legato sends its own computer function in OpenAI format, with Claude-style action names such as left_click and key.
You call Claude on Bedrock or Vertex. Legato needs an OpenAI-compatible endpoint.
You want a reference to read and modify, with a Streamlit chat and a VNC view of the desktop.
You want Anthropic's own fix for the same costs. The computer-use-best-practices quickstart (macOS only) adds a computer_batch tool and prunes screenshots at intervals so the cache holds [13]. We did not compare against it.

Code

pip install legatolabs
pip install "legatolabs[e2b]"   # E2B Desktop

import os
import legato
from e2b_desktop import Sandbox

agent = legato.Harness(model="anthropic/claude-sonnet-5.5", base_url="https://openrouter.ai/api/v1",
                       api_key=os.environ["OPENROUTER_API_KEY"])
desk = Sandbox.create()                     # needs E2B_API_KEY
try:
    result = agent.run("Add a Total row under the table and save", legato.attach(desk))
    print(result.report())                  # model, action, wait and screenshot time
finally:
    desk.kill()

legato.Harness.reference(...) takes the same arguments and runs the reference loop. To reproduce its timing, give it a desktop adapter that sleeps 2 s per action.

Sources

Anthropic links are to anthropics/claude-quickstarts at commit 4a4224bdf0e9cfb3332167b69c90457403813625.

Providers: computer_use_demo/loop.py#L124-L130
Container and single session: computer-use-demo/README.md#L34
Tool groups: computer_use_demo/tools/groups.py#L36-L63
Screenshot delay: tools/computer.py#L99 and #L273-L274
Image truncation: loop.py#L132-L139 and streamlit.py#L208-L209
Token budget, retries, tool loop: loop.py#L101, #L125, #L202-L239
computer-use-demo/LICENSE
Legato loop and defaults: legato/harness.py
Sandbox adapters: legato/env.py
Ready signal: legato/settle.py
Call timeout and retry: legato/rl/policy.py
Method, run H2H-1 (12 November 2026) and raw data: Benchmarks
Best-practices quickstart: computer-use-best-practices/README.md and constants.py#L149-L157

Edit this page on GitHub · Markdown