WebsiteBench
Can AI agents rebuild websites through browser exploration?
Computer-use agents, indexed.
Benchmarks, datasets, models and papers for agents that use computers, each checked against its primary source.
Star on GitHubStar Awesome CUA518
2,170 reviewed records1,673 papers and surveys2,062 scan finds not yet in the catalogScanned
More 86
Results
- Featured
-
Benchmark
WebsiteBenchTIGER-AI-Lab/WebsiteBench
Website reconstruction benchmark: agents explore a production-level site through a browser, rebuild a full-stack version, and are scored by hidden functional and visual test cases across 70 tasks.
WebReviewed -
Benchmark
ClawBenchTIGER-AI-Lab/ClawBench
Everyday-task benchmark on live websites that runs browser agents in containerized Chromium, blocks each final submission request for scoring by URL match and an LLM judge, and records five layers of run evidence.
WebReviewed - Most mentioned across the catalog
-
Benchmark
OSWorldxlang-ai/OSWorld
Execution-based evaluation of agents performing tasks in real desktop environments.
DesktopReviewed -
Benchmark
WebArenaweb-arena-x/webarena
Self-hosted web environments and tasks for evaluating autonomous web agents.
WebReviewed
Benchmarks
Featured benchmarks, then every reviewed record that publishes an official leaderboard.
ClawBench
An everyday-task benchmark for browser agents on live websites.
Official leaderboards
20 records, sorted by catalog mentions
- OSWorldDesktop
- WebArenaWeb
- ScreenSpot ProDesktop
- AndroidWorldMobile
- AndroidLabMobile
- WorkArenaWeb
- MobileWorld (Tongyi-MAI)Mobile
- Online Mind2WebWeb
- AgentBenchWeb and 1 more
- MacAgentBenchDesktop
- TheAgentCompanyWeb and 1 more
- AgentRewardBenchWeb
- AppWorld
- ClawBenchWeb
- Mind2Web 2Web
- RecreationBenchCross-platform
- SaaS-BenchWeb
- SafeArenaWeb
- WebCanvasWeb
- WebsiteBenchWeb
Start from a task
Each one opens the catalog already filtered for the job.
91 BenchmarksEvaluate an agentOSWorld, WebArena, AndroidWorld86 DatasetsTrain on GUI trajectoriesGUI-Odyssey, GUIAct, AndroidControl98 ModelsGround clicks on screenSeeClick, UI-TARS, OS-Atlas87 FrameworksBuild or run an agentBrowserGym, Agent S, OpenCUA48 Tools and infrastructureSandbox agents safelyE2B, Cua, AndroidEnv289 Safety researchCheck safety and securitySafeArena, AgentDAM, RiOSWorld
Latest
Newest review batch: 2026-11-06. Scans run every hour; their finds stay labelled until reviewed.
Recently reviewed resources
- web-eval-agentReviewed
- TestZeus HerculesReviewed
- TestDriver SDKReviewed
- TactileReviewed
From the latest scans UTC
- Cerebrum (AIOS SDK): A Platform for Agent Development, Deployment, Distribution, and…Metadata only
- From Google Gemini to OpenAI Q* (Q-Star): A Survey of Reshaping the Generative Artificial…Metadata only
- Timeline-Bench: Evaluating Agents on Realistic Video-Editing Tasks, from Raw Footage to…Metadata only
- FDE-Bench: Evaluating LLM Agents for Deployment Environment ConfigurationMetadata only
Browse by category
- Papers1,645
- Benchmarks91
- Models98
- Datasets86
- Frameworks87
- Agents49
- Surveys28
- Tools32
- Infrastructure16
- Products20
- Publishers16
- Learning2
Papers
1,673 reviewed papers and surveys, newest first. Bibliographic details are checked against arXiv or the publisher.
For agents and developers
The whole catalog is open data. Pull it into a notebook, a retrieval index or your own agent; every record keeps its evidence links and review state.
# Every record, one JSON object per line
curl -s https://awesomecua.com/downloads/catalog.jsonl \
| jq -r 'select(.kind=="benchmark" and .review_status=="curated") | .name'from datasets import load_dataset
ds = load_dataset("TIGER-Lab/AwesomeCUA", "resources", split="train")
benchmarks = ds.filter(lambda r: r["kind"] == "benchmark")
print(benchmarks[0]["name"], benchmarks[0]["url"])# An index of the site for language models
curl -s https://awesomecua.com/llms.txt@misc{awesomecua,
title = {Awesome CUA: a source-backed catalog of computer-use agent resources},
author = {{Awesome CUA}},
url = {https://awesomecua.com}
}Updated with every scan. Licence: MIT.
Computer-use agent research and resources
Browse source-backed pages directly, with or without JavaScript. Reviewed entries include concise annotations, provenance, and separate license declarations.
中文指南Know a resource that belongs here?
A reviewer reads the primary source first, and the record links to it.