AIエージェント
open-codex-computer-use
Open-source Computer Use as an MCP service: let any agent drive macOS, Linux, and Windows desktops.
★ 2.4k スターMIT
ガイドを見る →agent-desktopは、lahfirが公開しているオープンソースのAIエージェントプロジェクトです（GitHub: lahfir/agent-desktop）。GitHubスターは1,778、フォークは127で、Apache-2.0ライセンスの下で公開されています。GitHubでの最終更新は2026年10月6日です。
Reliable computer use for any agent: drive real apps through accessibility trees, not pixel guessing.
Decide which desktop workflow your agent must perform and which apps it involves.
Install agent-desktop globally from npm, or build it from source with cargo.
Grant macOS Accessibility permission (and Screen Recording if you need screenshots), using the built-in permission request command.
Take a snapshot of the target app, note element references, act on them, and re-snapshot to verify the result, optionally inside a traced session.
Snapshots return compact element references tied to a snapshot id, so agents act on identified controls rather than screen coordinates.
Dense apps can be explored through a shallow skeleton overview with targeted drill-down, which the project says cuts snapshot token use dramatically.
Reference-based actions use accessibility APIs and avoid silently stealing focus, cursor, keyboard, or clipboard.
Dozens of commands cover observation, interaction, keyboard, mouse, notifications, clipboard, window management, sessions, and trace export, all with structured JSON output.
Launching a Chromium app with CDP opens a verified DevTools port so Playwright, Puppeteer, or agent-browser can drive web content while native UI stays on the accessibility path.
Shared sessions give multi-agent workflows a common snapshot namespace and JSONL traces exportable as a single HTML viewer, and a C-ABI library supports in-process use from many languages.
The npm install is the recommended route; source builds need Rust 1.89+ and macOS 13+.
git clone https://github.com/lahfir/agent-desktop.git
cd agent-desktop
cargo build --releaseThe package downloads the prebuilt binary automatically.
npm install -g agent-desktop
agent-desktop permissions --requestList a Finder window's interactive elements, then click one by reference.
agent-desktop snapshot --app Finder -i
agent-desktop click @e3 --snapshot <snapshot_id>It is maintained by lahfir on GitHub.
Yes, it is Apache-2.0 licensed and free; the AI model that decides what to click comes from your own agent setup.
Agent Desktop needs no API key; it is a local control tool. The agent driving it uses your existing model access.
Screenshot-based computer use guesses coordinates from pixels and breaks when layouts shift; Agent Desktop reads the real accessibility tree and acts on stable element references that are safe to retry.
Run npm install -g agent-desktop, grant Accessibility permission, then take your first snapshot with agent-desktop snapshot --app Finder -i.
Open-source Computer Use as an MCP service: let any agent drive macOS, Linux, and Windows desktops.
ESP32 iPhone Agent — let an AI model use a real iPhone with a $7 ESP32: AirPlay eyes, USB HID hands, no computer, no jailbreak
Benchmark computer-use agents on long, real-world desktop tasks.
AI agents that use real Android and iOS apps like a human — from minitap.
Pinvou Agent — open-source desktop AI agent for tools, files, knowledge, workflows and real deliverables
Jev Ultrafast — the fastest browser agent from browser-use: one decision-model request per action cycle drives a real browser, MIT
LibreAIカタログに掲載されたAIエージェントプロジェクトの中で最もスターの多いもの — agent-desktopが合わなかった場合に検討したい、現実的な候補リストです。
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
カタログの中でagent-desktopとスター数が近いAIエージェントプロジェクト — 前後の棚もそのまま見て回れます。
Dev agent multiplexer designed for git worktrees and coding agents.
Retired agent framework: its prompts and guides live on in successor repos.
Self-hosted ai agents: minimal ai coding agent (~1,000 lines of python) inspir.
Self-improving memory layer that works across AI agent harnesses.
Android GUI agent framework: AI that sees, plans, and taps real mobile apps.
Team memory for engineers and their AI agents.