update
May 26, 2026
By Teun
Vision Clicker auto-approves AI agent prompts on macOS
A new macOS app called Vision Clicker watches a selected screen region and automatically clicks approval buttons such as Run, Fetch, or Retry using Apple Vision OCR. The tool runs locally, needs Accessibility and Screen Recording permissions, and is intended for small approval prompts in coding-agent UIs.
Vision Clicker is a local macOS menu bar app that automatically clicks approval buttons in other applications. The project, published on GitHub as SlopeAutoAcceptor, is aimed at AI agent workflows where tools stop for a user confirmation such as Run, Fetch, or Retry.
According to the project description, Vision Clicker watches a user-selected area of the screen, uses Apple Vision OCR to read visible text, and then clicks the matching button. After the click, it restores the cursor to its original position. The app is designed for small approval controls, not for broad automation of an entire application.
The author says the app runs entirely on the local Mac and does not require an API key. It also does not download a model or send captured images to a server, according to the project notes. Settings are stored locally in UserDefaults, including the selected region, target labels, scan interval, and confidence threshold.
The app requires macOS 13 or newer, plus Accessibility permission to perform the synthetic mouse click and Screen Recording permission to capture the selected region. The project notes say the setup is tested with a dual-monitor configuration, including displays positioned above or beside the main screen.
To use it, a user clones the repository, runs the install script, and the installer builds the app, copies it into /Applications/Vision Clicker.app, reveals it in Finder, and launches it. If someone only runs scripts/build_app.sh, the app stays in dist/ and is not installed into Applications.
The workflow starts by drawing a capture rectangle similar to the macOS screenshot shortcut, then highlighting the saved region before running. Users can enter one or more labels, such as Run or Run, Fetch, and set a minimum confidence threshold. The project recommends 0.20 as a practical starting point for small buttons.
Vision Clicker can run once manually or keep scanning in Live mode. It also supports a Cursor-specific mode that sweeps across tabs with Cmd+Shift+] , clicks each visible target, and then returns with Cmd+Shift+[ . The project says the OCR matching is intentionally fuzzy after light normalization, so a label like Run can match text such as Running or Auto-Run.
The author warns that the tool should be used at the user’s own risk. The project disclaimer says AI agents can make mistakes and that company approval prompts often exist for real safety, privacy, compliance, and operational reasons. The app is meant to automate clicking only where the user has decided that behavior is acceptable.
For release workflows, the project includes a GitHub-backed release script that reads the latest release with gh, builds the app with the next patch version by default, pushes a tag, creates a GitHub Release, and uploads a zipped macOS app. Larger version bumps can be requested explicitly with an environment variable.
