typesafe-computer-use
macOS computer-use loop that OCRs the screen, has Jev pick the next action from a short list, and clicks, at about $0.0002 a step.
Overview
typesafe-computer-use drives a Mac toward a plain-English goal. It reads the screen deterministically with OCR and UI state, builds a list of candidate actions, and asks Jev which action comes next as a Choice with a calibrated confidence. A writing model is only called when a text field genuinely needs free text or when the classifier stops and the screen needs reading. Facts such as dates, focus state or whether a URL is clean are computed in code and passed as state.
The README compares one decision on the same screenshot and goal: Jev cost about $0.0002 versus $0.032 for Claude Opus 5 with a bare screenshot, with model latency of 0.13 to 0.38 s versus 5.2 s and about 1.5 s per end-to-end step. It notes the honest caveat that the larger model reads more. The project is in beta, drives the real mouse and keyboard, and recommends starting with a dry run.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate action should be taken next toward the goal? | — |
How it works
- Capture and OCR the screen; compute facts in code.
- Build candidate actions and ask Jev to choose.
- Gate on confidence; call a writer model only for free text.
- Click or type, then repeat until the goal is met.
Strengths
- About $0.0002 per decision in the README's measurement.
- Model latency of a few hundred milliseconds.
- No screenshots sent to a large model for most steps.
- Dry-run mode before acting.
Caveats
- Beta under heavy development.
- Drives the real mouse and keyboard.
- macOS only.
Getting started
clicker "<goal>" (start without --act for a dry run)At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- macOS CLI
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
macOS computer-use loop: OCR the screen, Jev classifies the next action, then click. About $0.0002/step.
AnotiaWang/awesome-decision-modelsBuilds candidate actions from OCR and UI state for Jev to control macOS, calling a text model only when writing is needed.
logicrw/awesome-jev-projectsmacOS computer-use experiment using OCR plus bounded Jev action selection.
cobanov/awesome-jevComputer use for about $0.0002 a step: OCR the screen, classify the next action with TypeSafe, click. macOS.
hellogumbo/awesome-jevmacOS computer use without sending screenshots to a large model. The screen is read deterministically and Jev picks the next action.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
Which action from a candidate list to take next on the screen.
How much cheaper is it than a frontier model?
The README measured about 155 times cheaper per decision than Claude Opus 5 with a screenshot.
Is it safe to run?
It controls the real mouse and keyboard; the author recommends a dry run first.
Related projects
Lets coding agents run whole goals in your signed-in Chrome on macOS in background tabs, asking Jev or Claude Haiku only for real choices.
ProjectbetaComputer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonLocal stdio MCP server that executes whole browser and macOS goals with a Jev-driven observe-decide-act loop.
MCP serverbetaTypeScriptBrowser Use's browser agent where Jev picks an operation and page element in one request and a small LLM only writes text to type.
LibrarybetaPythonHelpers for browser-harness that let Claude Code or Codex hand off a whole website sub-task in one call, with Jev deciding each step.
ProjectbetaPythonPython decision layer for browser and desktop automation: code observes, acts and verifies; the model makes one typed choice per step.
ProjectbetaPython