agent-desktop
CLI for agents to operate apps through the accessibility tree
Worth exploring
After each action, read the interface again and use identifiers from the new snapshot
Read, not run
What it can touch
- Node.js
- Open code on GitHub · Apache-2.0 license
- Nov 2026 · by @vercel-labs
We read the code and docs on this date. We did not run it. Check before you install.
How we checkOn 18 Nov 2026 we read the code and docs on GitHub. We found that it runs commands in your terminal. We did not run it.On 18 Nov 2026 we read the code and docs on GitHub to see what you need before it works. We did not run it.On 18 Nov 2026 we checked the GitHub repository and its license.This comes from the GitHub repository, as we saw it on 18 Nov 2026.Wunderlust read the code and docs on GitHub for this card on 18 Nov 2026. We did not run it. It is not a security audit. Check before you install.
What it is
It reads macOS, Windows, and Linux apps and returns elements as JSON with action identifiers. The observe, act, and verify loop lives in the agent, with no model inside the tool.
What it’s for
Agent developers can use snapshots and identifiers to click or type in desktop apps.
Benefits
- Actions on accessibility tree elements
- CLI, C library, and MCP server
Caveats
- Identifiers belong to a snapshot and need to be renewed when the interface changes.





