An archive you can search by how something looks, and move through by why one thing leads to another.
Music video, 1966 to 2026
Match Cut is a music video archive that you and your agent are meant to use at the same time, on the same screen.
Every video here carries a measured visual fingerprint: motion, brightness, warmth, saturation, contrast, average shot length. You could not put that behind a filter panel anyone would use. Six continuous dimensions, plus a request like "something shot like this one but from a different decade", is not a control. It is a sentence. The tool find_by_look exists because language is the only sensible interface to that question.
The archive holds 49,753 links between videos, and not one of them is a similarity score. Each states the reason it exists: a shared director, era, movement or tag. So when the agent moves sideways through the archive it can tell you exactly what it followed, and you are free to disagree with it.
Turn on Keep it going and the agent takes over the screen: when a video ends it chooses the next one and says in a line why it went there, usually by following a connection so the run has a thread through it. It is a station programmed live, with its reasoning out loud, and it is off until you ask for it.
The page offers eight tools through WebMCP, and each one maps to something you can also do by hand. Every call is printed in the conversation with its arguments and what came back, so you can see how the agent got somewhere rather than only where it landed.
Each video carries cultural context, a curatorial assessment and a director biography, and every attribution carries a recorded confidence: confirmed, likely or unknown. Where the archive is unsure, it says so on screen rather than quietly rounding up, and the agent is told to flag that rather than state it flatly. An archive that tells you how sure it is beats one that sounds certain about everything.
Match Cut is not a standing start. It sits on two projects by the same builder.
MTV Rewind is a continuously running music video station: roughly 100,000 videos across 87 curated channels, programmed rather than recommended, with a synchronised room where people watch the same thing at the same time. The premise underneath this page, that watching together beats a list of results, was worked out there first.
RAD, the Rewind Archive of Data, is the research layer: the 7,139 annotated records this archive runs on, spanning 1966 to 2026, with per-label archives, a director index and the connection graph. The cataloguing behind it is rigorous and academic, and none of it was reachable by an agent until now. WebMCP is what made it addressable in a sentence.
The corpus and the annotations are used here with the permission of their author.
Ask your agent, or press a line to run it here