Show HN: ProofShot – Give AI coding agents eyes to verify the UI they build
TL;DR
ProofShot is a CLI tool that gives AI coding agents (Claude Code, Cursor, Codex, etc.) actual browser vision – they can open pages, click around, take screenshots, and capture console errors.
Key Points
- The agent records a session via shell commands and bundles video, screenshots, and logs into a single self-contained HTML file for quick review.
- Built on Vercel Labs' agent-browser, reportedly much faster than Playwright MCP – not a testing framework but a verification layer.
- Ships as a 'skill' so the agent understands its own capabilities and can use them deliberately.
Nauti's Take
The concept is simple and clever precisely because of that: don't make the model smarter, fix the information flow. Anyone who's watched an agent proudly commit a broken layout because it never saw a screenshot understands the pain point immediately.
Basing it on shell commands and keeping it agent-agnostic is the right architectural call – no vendor lock-in, maximum compatibility. Open questions remain around complex SPAs or auth-walled apps, but as a direction, ProofShot signals where things are heading: agents need senses, not just more parameters.
Context
AI agents write enormous amounts of UI code today but are essentially blind to what actually renders in the browser – a fundamental gap that causes silent layout breaks and ignored JS errors. ProofShot closes this feedback loop without heavy test infrastructure: the agent sees what it built and can self-correct immediately. This isn't a nice-to-have; missing visual verification is one of the top reasons vibe-coding sessions devolve into manual bug-fixing rounds.
Sources