🤓 Latest Submissions
OpenLook
OpenLook is a video-first visual unit testing MCP for coding agents. Today, AI coding agents can write UI code, run type checks, execute unit tests, and use Playwright to verify that DOM elements exist. But they still cannot reliably tell whether the interface actually looks clear, polished, usable, or aligned with product intent. OpenLook closes that feedback loop. Developers write simple `.openlook/*.yaml` specs that describe a URL, browser steps, and visual checks. A coding agent uses Playwright MCP to record the real browser session as a WebM video, then sends that recording to OpenLook MCP. OpenLook passes the recording and checks to Gemini, which evaluates the actual user experience: layout, visual hierarchy, motion, contrast, clarity, and task flow. The result is a structured visual audit with pass/fail checks, reasoning, reports, and recommended fixes. The agent can then patch the UI and rerun the same visual spec until it passes. This turns subjective UX review into a repeatable agent workflow. For the IBM Bob Hackathon, OpenLook demonstrates Bob acting as a full development partner: running visual tests, interpreting failures, fixing a deliberate UI regression, and verifying the repair through the same OpenLook spec. Works better with stronger reasoning models.
17 May 2026
