AgentScreenshots

About AgentScreenshots

AgentScreenshots closes the loop that makes AI coding agents bad at front-end work: they write CSS and rebuild component trees without ever seeing the rendered pixels, so they finish by asking you whether it looks right. The tool is a single CLI command — agentshot, pointed at a URL with a selector — that renders the page in a local headless browser and drops a PNG on disk in roughly two seconds. Because it captures a scoped section rather than a whole page, the agent gets exactly the component it just edited instead of a wide-angle screenshot that floods its context, and because it runs as one shell call rather than a browser MCP session, there are no navigation events or DOM dumps in the transcript. Localhost is a first-class target: no tunnel, no proxy and no public URL are needed, so a Vite dev server on port 5173 is captured directly. Selectors can be CSS, text or XPath with nth-match and vertical slices, any viewport width from mobile to ultrawide is supported, and it can click or hover first to dismiss banners, open menus and capture prepared states, and scroll to trigger lazy-loaded images. Output is retina by default with a 1× mode for large batches, --json emits metadata for pipelines, and a drop-in agent instruction package teaches Claude Code, Codex, Cursor, Windsurf or OpenCode when and how to capture.

Key Features

Tags