telexed ~ c / 063f6127-a6fradar:60 · agent_toolLIVE
← back
NO.
#063f6127
Topic
AGENTS & TOOLS
Source
Hacker News · Show HN AI
Published
2026-05-02 02:18:24
Importance
★ 6/10 — radar 60
`agent-desktop`: structured native desktop automation CLI for AI agents
FIG-0631:1

`agent-desktop`: structured native desktop automation CLI for AI agents

Desktop agents can skip pixel-click loops and operate on the accessibility tree instead. That cuts token cost by 78% to 96% in heavy apps and makes local app automation practical enough to test now.

[ KEY POINTS ]
  1. Instead of screenshot -> coordinate prediction -> click, it uses OS accessibility APIs on macOS, Windows, and Linux, so actions target real UI elements.
  2. The CLI ships as a single ~15 MB Rust binary with 53 commands and JSON output, making it easy to wire into agent loops and scripts.
  3. Progressive skeleton traversal avoids dumping the full UI tree; Slack-sized trees can exceed 50,000 tokens, so shallow snapshots plus subtree fetches matter.
  4. Element refs like @e12 and subtree re-querying let agents act, then refresh only the changed region, which is faster than full re-snapshots.
Originalgithub.com/lahfir/agent-desktopRead original →

// related