What does an AI agent actually see when it searches your footage?
Not the video. When an agent searches a local index, the heavy work happens on your machine and the agent receives text: file paths, time ranges, a one-line description of each shot and sometimes a line of dialogue. That text does go to the agent's model provider, so the honest answer for NDA work is "less than an upload, more than nothing".
Published October 15, 2026 · Field notes
"100% local" is an easy claim when you are the only one asking the questions. Once an AI agent does the asking, it is fair to ask what that agent gets to see. Most agents - Claude, Cursor, Codex - send everything they read to a model running in someone else's data centre. So the useful question is not "is the search local" but "what does the agent read".
Three things that never leave
- The video and photo files. They are read from disk by the app on your machine.
- The index. Embeddings, transcripts, face data and GPS stay in the local database.
- The rendering. MP4s and FCPXML files are written on your machine, to a path on your disk.
What the agent receives
Using Reelary as the example (it is our app), each search the agent runs returns a short text list. One line looks roughly like this:
3. /Volumes/Archive/2025/Lisbon/DJI_0412.MP4 00:14-00:20 (start=14.00, end=20.00) score=0.312 match=description shows: aerial view of waves breaking on rocks below a lighthouse
So the agent, and its model provider, learn:
- File paths. Folder names often carry client and project names. This is usually the most revealing part.
- Time ranges and scores. Harmless on their own.
- A one-line description of the shot, written by a vision model on your machine.
- The matching line of dialogue, when a Pro dialogue search found the clip by what was said.
- Names you gave people, if the agent lists people or their clips (Pro), and place names from GPS (Pro).
- Small preview frames, only on request. The search tool can attach up to eight thumbnails when the agent asks; by default it does not.
Compare that with an upload: hours of transfer and a full copy of the material on a third-party server. The text an agent sees is closer to pasting a folder listing with notes into a chat.
Is that acceptable for client work?
It depends on the agreement, and you should read yours. Many NDAs are about the material itself, which never moves.
Some cover any information about the project, and a path like /Clients/BrandX/Unreleased_Launch plus a
line of dialogue arguably counts. Check what your model provider does with API or chat data - retention and training
terms differ between consumer, team and enterprise plans - and decide per project. Our wider take on the upload
question is in does searching your footage upload it.
How to keep a project fully local
- Use the app instead of the agent for that project. Same search, same editor, no model provider involved.
- Use a local model. An MCP client pointed at a model running on your own machine keeps the text local too.
- Keep sensitive projects out of the indexed folders the agent can search, or remove a folder from the library while you work with an agent.
- Ask for no thumbnails. Tell the agent to work from descriptions only.
The short version.
With a local index, an agent sees a list of matching clips, not your footage. That list is still information, and it goes wherever your agent sends its context. For most libraries that is fine; for the ones where it is not, the app works the same way without the agent. Setup and the full tool list are on the Reelary for AI agents page.
Questions people ask about this
Is using an AI agent with my footage the same as uploading it?
No. An upload sends the video. An agent calling a local index receives a text summary of search results: paths, times, short descriptions. Nothing is transferred that would let someone play or reconstruct the footage, unless preview frames are requested.
Can I stop the agent from receiving preview images?
In Reelary, thumbnails are only attached when the agent sets an option on a search. Most searches do not. If you want none at all, tell the agent not to request thumbnails, or use the app directly.
Does Reelary send anything itself?
No. The app makes no requests with your library data. Anything that leaves the machine in an agent workflow is what the agent itself forwards to its model provider.
How do I keep an NDA project fully offline?
Use the Reelary app without an agent, or an agent connected to a model running on your own machine. Either way no text about the footage leaves the computer.