Vercel Labs' open-source command-line tool and skill that let your agent read and operate native desktop apps through the system's accessibility tree: it gets a numbered list of real buttons and fields, acts on one, then looks again to check, with no guessing from screenshots.
In short: Vercel Labs' open-source command-line tool and skill that let your agent read and operate native desktop apps through the system's accessibility tree: it gets a numbered list of real buttons and fields, acts on one, then looks again to check, with no guessing from screenshots. It is a free skill for people who use Claude Code, Codex or Cursor and want help with coding agent skills.
Checked on · Added
| Agent / tool | Access / cost | Link |
|---|---|---|
| Claude Code | Paid Claude plan or API | claude.com/claude-code |
| Codex | ChatGPT plan or API | openai.com/codex |
| Cursor | Free tier + paid | cursor.com |
| GitHub repo | Open source · ★ 11.4k | vercel-labs/agent-desktop |
Image: from the GitHub README, shown with credit.
npm install -g agent-desktop npx skills add https://github.com/vercel-labs/agent-desktop --skill agent-desktop agent-desktop permissions --guide
Install command copied from the project README. Check the README for the latest version.
Vercel Labs' open-source command-line tool and skill that let your agent read and operate native desktop apps through the system's accessibility tree: it gets a numbered list of real buttons and fields, acts on one, then looks again to check, with no guessing from screenshots.
agent-desktop is Vercel Labs' answer to computer use without screenshots. Instead of guessing where to click from pixels, it reads the accessibility tree that screen readers use: a snapshot returns JSON with a reference…
Short excerpt from the README of vercel-labs/agent-desktop. Read the full README on GitHub.
Written by us so you can reuse it. Replace each [bracket] with your details.
Snapshots, screenshots and clipboard reads can contain whatever is on your screen, so keep them out of public logs. It only sees what apps expose to accessibility tools: canvases, games and custom-drawn apps stay blind spots. On Linux, mouse input and window control need X11.
Vercel Labs' open-source command-line tool and skill that let your agent read and operate native desktop apps through the system's accessibility tree: it gets a numbered list of real buttons and fields, acts on one, then looks again to check, with no guessing from screenshots.
It works with Claude Code, Codex and Cursor. The code is on GitHub at vercel-labs/agent-desktop.
Free & open source (Apache-2.0). No account or API key for the CLI. Runs with the agent plan you already pay for. The optional JEV helper sends screen descriptions to a model provider you configure.
Run the install command from the README: “npm install -g agent-desktop”. Check the README on GitHub for the latest version, then follow the steps on this page.
We read 486 instruction and code files in the repo and found no red flags. That is not a guarantee, so review anything that runs commands on your machine. How to check a skill before you install it →
All credit to the original authors. We rewrite in our own words and link back. Something wrong? Report it.
SKILLYour coding agent connects to an Android or Google TV over Wi-Fi, switches off the ad and tracking apps, and puts in a clean, Apple TV-style home screen. Nothing is uninstalled.
Claude Code +2★ 14.8k
SKILLA pack of 14 skills from a design engineer that make your agent review and polish interfaces with exact rules: ranked UI reviews, typography, colour, accessibility, layout and product writing, plus tools that stress-test components and draw technical illustrations.
Claude Code +2★ 10.3k
SKILLA Claude Code skill that records an animation from a live site, a screen recording, a gif or a Lottie file, rebuilds it in your own code, then films the copy the same way and compares the two, round after round, until timing and motion match.
Claude Code★ 809
SKILLA Claude skill that reviews a vibecoded screen like a designer: it lists the tell-tale AI touches, proposes a direction from your own product, then fixes spacing, colour, icons, copy, type and layout pass by pass. Works on dashboards, apps, landing pages, or a live site from a link.
Claude Code★ 540
SKILLOne install gives your agent a plan, test, build, review and remember routine, with 71 helper agents, about 300 skills, memory between sessions and a built-in scanner for risky agent settings.
Claude Code +3★ 285.3k
SKILLA skills framework that makes agents brainstorm, plan, test and debug systematically.
Claude Code +1★ 299.2kFree newsletter
New AI tools you can use right away, in plain words. Every Friday.