macOS Harness
A minimal open-source toolkit that gives an AI agent hands and eyes on your Mac — it can see the screen, move the mouse, type, and click, without needing a custom integration for every app.
🔗 Visit macOS Harness
Description
Most "AI controls your computer" tools build a specific integration for each application they support, which means new apps are always unsupported until someone builds a connector. macOS Harness takes the opposite approach: it gives an LLM six general-purpose primitives — see, key, type, click, ax (accessibility), and script — that work across any macOS app, the way giving someone a keyboard and mouse works for any program rather than needing app-specific remote controls.
From the team behind browser-use (a well-known browser automation platform), it supports background window capture without needing to bring an app to the foreground, can target processes directly by PID, and shows an animated click-through pointer overlay so you can watch what the agent is doing in real time. Python scripting, browser automation, and filesystem access all run in the same process, keeping the primitives simple to compose. It's explicitly marked as experimental software, is macOS-only, and includes anonymous telemetry by default (limited to CLI commands and success metrics, not prompts or screenshots) along with a 'macos-harness doctor' diagnostic command.
💬 Our review
The short version: as a minimal, general-purpose control layer rather than a polished product, macOS Harness is best understood as infrastructure for developers building their own computer-use agents on macOS, not something an end user installs directly.
Against Anthropic's Claude Computer Use, which is a comparable general-purpose "see and click" capability but delivered as part of a hosted model API, macOS Harness is the open-source, self-hosted, macOS-native alternative — you get more control and no per-call API dependency, at the cost of needing to wire it into your own agent loop yourself. Coming from the browser-use team lends it credibility, since they've already shipped a working automation platform, but "experimental" is an accurate label: primitives this low-level require careful prompting and error handling to use safely, and background/PID-targeted control on someone's real desktop is inherently higher-risk than sandboxed browser automation. Solid pick for developers building macOS computer-use agents who want full control; not something to point at a real Mac account without safeguards in place.
💰 Pricing
📊 Global score
🤖 AI-enriched data
100% gratuit, licence MIT.
Pros
Six primitives génériques (see, key, type, click, ax, script) fonctionnant sur n'importe quelle app
Capture de fenêtre en arrière-plan sans mise au premier plan
Ciblage direct par PID, overlay animé du curseur pour observer l'agent en action
Python, automatisation navigateur et accès fichiers dans le même process
Créé par l'équipe de browser-use, plateforme d'automatisation reconnue
Cons
Marqué explicitement expérimental — pas prêt pour un usage non supervisé
macOS uniquement, aucun support Windows/Linux
Télémétrie anonyme activée par défaut (désactivable)
Primitives bas niveau : nécessite du travail d'ingénierie prompt/sécurité pour un usage fiable