MCP for video libraries: connecting AI assistants to your footage
AI assistants are good at writing, planning and research, but they know nothing about the footage on your drives. The Model Context Protocol (MCP) is an open standard that lets an assistant call tools in other applications. A video library that speaks MCP can let an assistant search your recordings and cite what it finds.
Updated
What MCP is
The Model Context Protocol defines how an AI application (the client) discovers and calls tools offered by another program (the server). Instead of pasting context into a chat, the assistant asks the server for what it needs, such as a search, a document or a status check, and receives structured results.
What an assistant can do with a video library
Video files are large and opaque to a language model. The useful layer is the evidence extracted from them. A video library MCP server can let an assistant:
- Find moments that match a description and return their source and time range.
- Ask a question across selected recordings and receive an answer with citations.
- Read the transcript for a bounded time range.
- Check coverage: which videos are indexed and which drives are available.
That turns requests like “find the three best moments where the founder talks about pricing and draft a post around them” into work the assistant can actually ground in your footage.
Questions to ask before you connect one
| Question | What a good answer looks like |
|---|---|
| What can the assistant reach? | Only the folders and evidence categories you grant, per connection |
| Can it change anything? | No: read-only, with no deleting, moving, renaming or exporting |
| Do new folders join automatically? | Not unless you explicitly choose an all-future scope |
| Can I revoke access? | Yes, at any time; revocation stops all further requests and citation lookups |
| Is the server exposed to the network? | No; local transport with authentication |
| Can results be checked? | Every result carries a source, time range and evidence type |
Footage can contain instructions
Speech and on-screen text are untrusted input. A recording could contain someone saying “ignore your previous instructions”. A well-designed server returns transcripts as source data and never lets content inside a video change what a connection is permitted to do.
How CutroomOS does it
CutroomOS runs a local MCP server that stays off until you create a connection. Each connection has its own folder grant and evidence categories, access is read-only and revocable, and every result includes a stable source ID, time range and evidence type, plus a link to review the moment in the app. The tools include search_clips, ask_library and get_transcript. See Connect through MCP and the setup guide.