Jev.aitools.fyi

Command Palette

Search for a command to run...

Featured
Product profile
Rust1,413 starsUpdated 2026-09-21

agent-desktop

agent-desktop (github.

Overview

agent-desktop (github.com/lahfir/agent-desktop) is a Rust-native CLI distributed on npm. It snapshots any app's accessibility tree, returns compact skeleton overviews with drill-down refs, and executes headless-safe clicks, typing, scrolling, and window management. Your harness calls TypeSafe Jev to choose the next ref action from structured JSON, the same pattern Muhammad Aayan (@socialwithaayan) summarized as Jev picks the next button or input.

Problem: Desktop agents that read pixels or brittle DOM dumps waste tokens and mis-click when UI refs shift between snapshots.

Built for: Agent builders on macOS who want native accessibility-tree computer use with typed next-action routing.

First indexed on Jev Directory: 2026-09-21

Creator and team

Name
lahfir
Handle
@lahfir
“Desktop automation on the accessibility tree. Jev picks the next button or input.”

How Jev is used

Role in the product flow
Choice over the next desktop ref action in the observe-act loop
Primitives
Choice
State in
agent-desktop snapshot JSON: skeleton regions, qualified refs (@snapshot:eN), roles, names, and values from macOS accessibility APIs.
Decision out
Selected command family (click, type, scroll, etc.) and target ref from the current finite action menu.
  1. agent-desktop snapshot --skeleton (or find) returns refs and snapshot_id
  2. Harness builds allowed actions from refs and safety policy
  3. Jev Choice picks operation and ref in one System One call
  4. agent-desktop executes via accessibility APIs; loop until goal or budget

agent-desktop deliberately does not embed an LLM. It is the hands and eyes: structured observation in, deterministic CLI commands out. Jev sits in the calling agent exactly like Browser Use Ultrafast: rebuild the action space from fresh state, ask one typed question, threshold confidence, then act. Refs stay stable across steps because they are accessibility identities, not coordinates. Pair with Hermes, Claude Code, or custom harnesses via npm global install, npx, or the C-ABI cdylib for in-process calls. Chromium apps can mix CDP for web content with native AX for menus and dialogs. This profile does not claim a bundled Jev binary inside agent-desktop; wire console.typesafe.ai credentials in your agent layer.

Sourced performance claims

  • Progressive skeleton traversal reports seventy-eight to ninety-six percent token reduction on dense apps versus flat snapshots in README examples.Source: github.com/lahfir/agent-desktop README

Features and stack

Features

  • Rust CLI with fifty-eight command names and structured JSON errors
  • Skeleton snapshots with drill-down refs for Slack, Finder, Xcode, and more
  • npm and npx install with prebuilt macOS binaries
  • ClawHub and skills.sh agent-desktop skill docs
  • Featured in Aayan top-ten Jev repos thread

Stack

  • Rust
  • macOS Accessibility APIs
  • TypeSafe System One (in agent harness)

Pricing: Open source Apache-2.0; TypeSafe API usage is bring-your-own when your agent calls Jev.

FAQ

Does agent-desktop include Jev?
No. It exposes desktop actions. Your agent or skill harness calls Jev separately to pick the next safe action from snapshot output.
Which platforms are supported?
macOS thirteen plus is production-ready per README. Windows and Linux adapters are planned with the same core contracts.

Why accessibility beats pixels

Native AX trees expose roles, names, and stable refs that survive window moves. agent-desktop compresses dense apps with skeleton overview plus drill-down so Jev sees a small action menu instead of thirty thousand tokens of Slack chrome.

Related learn guides

Original Jev guidance that pairs with this product pattern.

  • Jev use cases

    The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.

  • Jev vs LLM classification

    When to gate with System One probabilities instead of asking a chat model to label things.

Related products

Hand-picked neighbors with rich profiles or overlapping tags.

Browser and computer usePython4,799
DemoFeatured
Jev picks the browser op and DOM target in one shot; a tiny LLM only shows up when something needs typing.

by @gregpr07

browser computer usebrowser-useagents
Browser and computer usePython828
Demo
Computer use for about $0.

by @awlevin

browser-computer-usebrowser-usemacos
Browser and computer useTypeScript78
jev-browser (github.
browser computer usebrowser-use
By Rishit