typesafe-computer-use
Drive a Mac using plain English goals via OCR and a TypeSafe classifier.
Drive your Mac with plain English goals for about a fiftieth of a cent per step.
typesafe-computer-use (jev) is an open-source tool that reads your Mac screen, classifies the next action using TypeSafe, and executes clicks for a fraction of a cent per step. It suits developers and power users wanting affordable, programmatic desktop automation without heavy frontier-model costs.
What it is
It combines Vision OCR, accessibility trees, and a small TypeSafe decision model to pick actions from a short list rather than planning every step with an expensive vision model. Code handles deterministic facts like date parsing and focus states, while a writing model is only called for free text or final answers.
Who it's for
It suits developers and tinkerers who want to experiment with desktop automation on macOS without paying steep frontier-model API fees. It does not suit users looking for a fully polished, zero-setup consumer application, as it requires terminal permissions, API keys, and careful dry runs.
Key features
TypeSafe Classification Loop
Uses a small classifier to make fast, cheap choices over action options instead of expensive frontier planning.
Deterministic Screen State
Combines Vision OCR and accessibility trees with code-computed facts like dates and URLs to inform decisions.
Offline Run Replay
Writes every run folder out so you can inspect timings, payloads, and replay stalled sessions offline.
Multiple Browser Backends
Supports both OCR screen captures and direct DOM perception over Chrome DevTools.
Up and running in 3 steps
Our honest verdict
What's great
- Significantly cheaper per decision than bare frontier vision models
- Fast model latency and end-to-end step times
- Maintains local run folders for offline debugging and replay
- Supports OSWorld benchmark evaluation out of the box
Where it falls short
- Requires active macOS Screen Recording and Accessibility permissions to function
- Depends on paid external API keys for TypeSafe and optional Anthropic features
- Under heavy beta development with changing behavior and experimental Windows support
Pricing & license
Free under its open-source license, but requires paid API keys for TypeSafe and optional Anthropic usage to drive the decisions.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
How much does it cost to run?
It costs about $0.0002 per step using the TypeSafe classifier, plus optional Anthropic API costs if you use the writer model.
What operating systems are supported?
macOS 14 or newer is fully supported, while Windows 10 and 11 are experimental.
How do I stop a runaway automation loop?
Press Ctrl-C in the terminal or slam your mouse into the top-left corner of the screen.
