Jamdesk has released an open-source skill that uses AI vision models to detect and blur sensitive data in screenshots, distributed as an installable package via `npx skills add jamdesk/skills --skill blur-image`. The company says the tool works with more than 30 coding assistants, including Claude Code, Cursor, Windsurf, Codex, and OpenCode.
The skill runs five sequential phases. First, an ImageMagick preflight check confirms the dependency is available. Then an AI vision scan extracts pixel-coordinate bounding boxes for sensitive regions—API keys, database connection strings, bearer tokens, customer email addresses—followed by a plain-English confirmation step where the user approves what gets redacted. ImageMagick's `-region WxH+X+Y` flag handles the actual pixel manipulation, and a final verification pass has the model re-read the processed image to confirm coverage. The AI layer does the classification work that pixel tools cannot: distinguishing a `sk_live_` Stripe key from a syntax-highlighted keyword, or spotting an Authorization header buried in curl output.
The release carries a specific security warning about blur sigma. Gaussian blur applied with a low sigma value can be partially reversed by deblurring algorithms—a meaningful risk given the high-contrast text typical of terminal screenshots. The author recommends sigma 20 or higher for real credentials, with solid black fill (`-fill black -colorize 100`) as the more reliable option when aesthetics are secondary.
The blur-image skill ships through the same skills CLI as Jamdesk's existing `update-jamdesk` tool for MDX documentation sync, both positioned as standalone agent capabilities that install into existing toolchains. The CLI is described as Vercel-backed, though that association is not confirmed by publicly available sources.