Use AgentVision
See and interact with the user's screen using screenshots, element discovery, clicks, form filling, and visual QA
/plugin install use-agentvision@rvanbaalenWhen to use
Use when Claude needs to see and drive what is on your real macOS screen: UI development feedback loops (before/after screenshots), visual QA of running apps, filling forms, navigating desktop applications, and testing iOS Simulator or Android emulators. Not for headless browser testing or Playwright/Puppeteer scripts.
Requires the agent-vision CLI (brew install rvanbaalen/tap/agent-vision) plus Screen Recording and Accessibility permissions.
How it works
- Opens a session with
agent-vision open <App>, which launches or activates the app and selects its window automatically — oragent-vision startfor a manual region drag - Scans the region with
agent-vision elementsto build an indexed list of accessibility and OCR elements - Acts on a discovered element — click, type, key, scroll, drag
- Waits for the UI to settle, then re-scans, because element indices go stale after every update
- Captures screenshots and reads them back to verify outcomes visually
- Stops the session once you confirm you’re done
Capabilities
Session management — every command is scoped to a session UUID covering a selected screen region. open handles app launch and window selection; start covers custom regions.
Element discovery — indexed scans over the Accessibility API and OCR. Complex apps return 80+ elements, so scans are filtered with jq down to interactive accessibility nodes.
Focus-free targeting — --element N drives the Accessibility API without moving the cursor or stealing focus, so you can keep working in another window. --at X,Y is the fallback for canvas and custom-drawn UI, and is always verified with agent-vision preview first.
Screen capture — screenshots of the region on demand, read back and described for before/after diffing and QA reports.
File sharing — attachments go in via an osascript clipboard copy plus Cmd+V, not by driving Finder dialogs.
Background delegation — long multi-step flows are dispatched to a background sonnet subagent that echoes one-line progress markers, streamed back through the Monitor tool while the main thread stays free.
Ground rules — never resizes or rearranges windows, never works outside the selected area, never sends blind keystrokes, and verifies outcomes on screen rather than through shell commands.
Invoke
/use-agentvision:use-agentvision
The skill also activates automatically when you say things like “look at my screen”, “take a screenshot”, “fill this form”, “check the UI”, or “I have X open”.