AI Screen Vision

Key things to consider before installing: - Verify the vision API endpoint and provider: the default SV_VISION_BASE_URL is https://api.gpt.ge/v1 (not an obvious official host). If you plan to use a cloud provider, set the base URL and API key to a provider you trust (OpenAI/Azure/Google) or use a local model. - Treat screenshots as sensitive: the skill base64-encodes and uploads full-screen images to the configured API. Only use with a trusted vision endpoint and avoid running on machines with sensitive information on screen. - Review and modify VNC/noVNC behavior: headless setup creates a VNC server and a noVNC web front-end and sets a default VNC password in install scripts. Change any default password, restrict access (bind to localhost or use firewall rules), or skip the headless/noVNC setup if you don't need remote web access. - Review system changes and run in isolation: the installer uses sudo to install packages and writes to /usr/local/bin and ~/.vnc. Consider testing in a disposable VM/container with restricted network access first. - Confirm safety policy limits: the skill includes safety_check logic, but it uses some hard-coded patterns and limits; review it to ensure it blocks actions you consider dangerous (customize confirm_before/blocked patterns as needed). - Metadata mismatch: registry metadata lists no required env vars, but the code uses SV_VISION_API_KEY and config.json; ensure the platform supplies those securely and that you understand where secrets are stored. If you decide to proceed: run the setup manually and inspect the scripts (setup scripts, sv-start content, config.json) before executing; avoid the headless/noVNC path unless you harden access and change the default password; and prefer a local vision model or a vetted cloud endpoint for image uploads.