๐Ÿฆž KinClawยทv1.8.0ยทMITยทmacOS only

The agent that operates
your real Mac.

A floating chat window that talks to a real LLM brain and drives your actual macOS desktop. Not a virtual sandbox, not a cloud VM. Open source, single 17 MB binary, zero subscriptions.

Run it in 30 seconds

Go 1.22+ required. Grant Screen Recording + Accessibility on first launch.

Install

$ go install github.com/LocalKinAI/kinclaw/cmd/kinclaw@latest

Launch the floating chat UI

$ kinclaw serve

Opens at http://127.0.0.1:8020/. For an app-like floating window: open -na "Google Chrome" --args --app="http://127.0.0.1:8020/?compact".

What makes it different

Cloud agents (Manus, Anthropic Computer Use, OpenAI Operator) run in their virtual machines. You can't see your Cursor, your Reminders, your private files. KinClaw runs on your Mac and operates the apps you already use.

Floating chat UI

kinclaw serve opens a single compact chat window in your browser. Operates your real macOS desktop alongside it โ€” not a virtual sandbox, not a remote desktop. Markdown streams, soul switcher, session JSONL recording + replay.

Voice in / out

๐ŸŽ™ push-to-talk via SenseVoice (CJK + English). ๐Ÿ”Š auto-spoken replies via Kokoro. Local audio service, no cloud roundtrip.

Self-forging skills

Hit a task with no skill? Pilot drafts SKILL.md, generates the script, registers it, retries the task. Circuit breaker prevents runaway loops.

Sub-agent dispatch

spawn skill forks a specialist child (researcher / eye / critic / coder / curator) on a different brain. Hierarchical, kernel-capped at depth 1.

The 5 claws

Each claw is independently usable. Four of them ship as pure-Go libraries you can pull into any project that needs to drive macOS at the framework level โ€” no Xcode, no cgo, no Swift bridge.

ScreenCaptureKit-based screen reading at sub-20ms.

Powered by sckit-go.

Semantic UI control via macOS Accessibility tree (AX).

94% of macOS apps. ui find / read / click / tree / watch.

Real mouse + keyboard via CGEvent.

Smooth motion, target_pid background mode (no focus steal).

Screen + audio capture, native MP4.

Non-blocking; tts narration captured in the same track.

web

Drive any web page via Playwright.

JS evaluation, click + type, screenshots, React-compat fill.

See the Embedded Dylib paper for the distribution pattern these libraries share.

How it compares

Other agent projects either run in the cloud (you can't see your real machine), or are bound to one app, or only operate browsers. KinClaw operates your whole macOS.

ProjectWhere it runsWhat it operatesSource
KinClawYour MacYour real macOS appsMIT ยท open
Anthropic Computer UseCloud Linux VMA virtual desktopClosed ยท API
OpenAI OperatorCloud browserA browser onlyClosed ยท beta
ManusCloud VMTheir virtual desktopClosed ยท subscription
Open InterpreterLocal terminalShell only ยท no GUIMIT ยท open
Cursor / Cline / AiderInside the IDEYour code onlyMixed

Part of the LocalKin family ยท macOS-only ยท v0.2 will ship a native Swift WKWebView shell with always-on-top

Star on GitHub