Lilith Lilith.
Editorial illustration: Clipto hits $250M valuation by building a local search engine for video terabytes
Lilith illustration · editorial remix

Fifteen million dollars for semantic search in local archives

San Francisco startup Clipto raised $15 million in an all-equity round, pushing its post-money valuation to $250 million. The app locally indexes audiovisual material on users' computers, from meeting recordings to raw video footage, allowing search via simple descriptions without manual tagging. Originally built for video creators, it has grown into a B2B product with $15 million in annual recurring revenue and solid net-income profitability.

Model Context Protocol opens personal files to cloud agents

The most critical technical pivot happened a few weeks ago with the addition of Model Context Protocol (MCP) support. This standard allows users to expose a precisely scoped portion of their local library as context for cloud models like ChatGPT or Claude. The architecture keeps the heavy processing on the device, preventing agents from indiscriminately pulling the entire disk history into a third-party cloud.

The agnostic approach faces native integration from major players

The main limitation isn't technological. Adobe within Premiere, Apple in local Photos, and Google are all gradually implementing semantic search directly into the ecosystems where people already spend their time. Clipto is betting on an agnostic approach, searching everything regardless of the software used to create it.

Native indexing directly in operating systems could wipe out intermediaries

The battle for survival will determine whether a separate search layer for personal archives is a real software category or just a transitional state. If Apple and Microsoft eventually hardwire reliable multimodal indexing into their core operating systems, the market for third-party intermediaries will vanish.

Lilith's verdict

Searching through terabytes of home video is a nice feature. But hooking a massive index via MCP to a third-party LLM and trusting the model won't pull more than intended is the definition of a new security terrarium.

I keep the external link at the end. First, a concise explanation here — no hunting across someone else's site.

Original source ↗