GitHub radar
Mac computer-use agent: 155x cheaper than Claude Opus
This tool drives a Mac toward a plain-English goal by reading the screen locally, calling a big model only when something actually needs to be typed.
typesafe-computer-use drives a Mac toward a goal you type in plain English. Instead of sending a screenshot to a large model on every step and waiting seconds for a plan, it reads the screen locally using OCR and the accessibility tree, then asks a small fast classifier which action to take next. The big model runs only when a text field genuinely needs input. Per the README table, one decision costs $0.0002 vs $0.032 with Claude Opus 5 (155x cheaper), and a 12-step task runs $0.003 vs $0.40–$0.90 (130x to 300x cheaper). Each step completes in about 1.5 seconds instead of 5.5. Requires macOS 14+, Python 3.12+, a TypeSafe API key, and optionally an Anthropic key for text-field actions.
Why a vibe-coder should care
If you automate repetitive browser or desktop tasks through an AI agent, this cuts the cost by two orders of magnitude compared to frontier-model computer use. You get the same result — the agent clicks through the screen on your behalf — but for a fraction of a cent per task instead of dollars.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Deploy typesafe-computer-use from https://github.com/awlevin/typesafe-computer-use — follow the README, set TYPESAFE_API_KEY and ANTHROPIC_API_KEY in .env (ask me for keys), then run `clicker "open the browser and go to arsentev.ai" --act`
Runs on macOS 14 or newer — just Python and a TypeSafe API key
Open on GitHub▌ More finds