First question
What can it touch?
The practical difference is not only whether the agent can browse or click. It is which websites, files, accounts, and review steps sit inside its reach.
Theme
AI Resources
Computer-use agents for browsers and desktops
A focused map for agents, models, and tooling that point AI systems at browsers, desktops, or visual interfaces.
Computer-use tools can touch accounts, files, websites, and private data. Use this page to narrow what to explore next, then check permissions, terms, and review steps before trying anything important.
Questions to check
Start with the job or constraint that matters now.
These checks frame the source-linked Resources below. They do not rank products or cover every option.
Human review
Keep review close
Look for permissions, confirmations, logs, screenshots, and ways to stop or correct the agent before it changes something important.
Testing path
Start with low-risk tasks
Use throwaway tasks, test accounts, and manual review before connecting anything private, paid, or hard to undo.
Coverage and freshness
Newest LifeHubber addition included here: June 9, 2026
These groups are selective starting points, not a complete directory. The date reflects the newest included Resource’s LifeHubber added date, not a recheck of every linked source. Check the original source for current setup, terms, limits, privacy, access, costs, and behaviour.
Fresh in this topic
Newer Resources already included in this map
1Recently added Resources from the groups below.
Browser and desktop control
Projects that operate interfaces
9Use this group when the agent needs to browse, click, type, observe a screen, or automate a visual workflow.
Browser Use
browser-use/browser-use
Why it fits this starting point
Browser-first navigation, clicking, typing, and custom tools keep the comparison on website control rather than a broader desktop runtime.
Open Computer Use
iFurySt/open-codex-computer-use
Why it fits this starting point
An MCP service with setup paths across Windows, macOS, Linux, Codex, and other clients shows a computer-control layer that plugs into an existing MCP-capable agent.
Cua
trycua/cua
Why it fits this starting point
Cua pairs a background desktop driver and agent sandboxes with benchmark and SDK layers for isolated computer environments that also need an evaluation path.
Holo3.1
Hcompany/holo31
Why it fits this starting point
Model sizes from 0.8B to 35B-A3B and local quantized checkpoints let web, desktop, or mobile control be matched to available hardware.
UI-TARS Desktop
bytedance/UI-TARS-desktop
Why it fits this starting point
A desktop application with vision-language control and local or remote operation adds a visible GUI-agent workspace beyond a browser library.
Gemma 4 Browser Extension
nico-martin/gemma4-browser-extension
Why it fits this starting point
Running through Transformers.js and WebGPU keeps this browser-agent experiment on-device when extension constraints and page-local models shape the choice.
Page Agent
alibaba/page-agent
Why it fits this starting point
In-page JavaScript control and text-based DOM interaction show natural-language operation embedded directly into a web interface.
MolmoWeb
allenai/molmoweb
Why it fits this starting point
Multimodal navigation from natural-language instructions makes visual page understanding the main execution path.
Lightpanda Browser
lightpanda-io/browser
Why it fits this starting point
A headless browser built for automation brings infrastructure footprint and agent-oriented execution into the comparison.
Also in AI
Follow the next layer.
Keep the thread going with AI Guides for decision habits for messy AI choices, AI Access for free and low-cost ways to compare AI model access, AI Ballot for a clearer view of what readers are leaning toward.