Can Claude search the videos on your computer?
Not by itself: Claude cannot open your drives, and uploading a footage library to a chat is the wrong shape of solution. What works is giving the agent a tool. Claude Code, Claude Desktop, Cursor and Codex can call a local video index over MCP, ask for "drone shot of the coast at sunset" and get back file paths with exact start and end times, while the video never leaves your machine.
Published October 15, 2026 · Can it do this?
The question usually comes from someone who already uses Claude for the script and the shot list, and then switches to Finder to scrub through rushes for an hour. Couldn't the assistant that wrote "drone shot of the coast at sunset" also go and find one?
On its own: no
Claude has no view of your drives. In a chat it sees what you attach, and an attachment is a single file in a single
conversation, not a library you can query next month. Even in Claude Code, which can list folders and run commands,
a file name like DJI_0412.MP4 says nothing about what is in the shot. The agent could try to open each
file and pull frames, but that means minutes per clip and thousands of images sent to the model for one search. The
same three walls apply as for ChatGPT and your video files: no
access, no persistent index, and an upload that takes days.
With a tool: yes
Agents that support the Model Context Protocol can call programs on your computer. That changes the question from "can the model understand my footage" to "is there a tool that already has". A local footage index has: it ran a vision model over every frame once, keeps the result on your disk, and answers a description in well under a second. The agent only has to ask.
In practice the exchange looks like this. You type "find drone shots of the coast at sunset". The agent calls the search tool with that phrase. The tool returns a ranked list - file path, start and end in seconds, a one-line description of the shot - and the agent tells you which clips to look at, or goes on to cut them into a sequence.
Which Claude
- Claude Code and Claude Desktop: yes, both run local MCP servers.
- claude.ai in a browser: no, a website cannot start programs on your machine.
- Cursor, Codex and other MCP clients: yes, the same server works.
Setting it up with Reelary
Reelary is our app, so read this part as disclosure. One command does the setup:
$ npx reelary It installs the app (or keeps the one you have), starts it in the background, registers the MCP server with Claude Code, Claude Desktop, Cursor and Codex, and adds a set of skills that teach the agent when to search, when to build a rough cut and when to hand off to an editor. Then you ask Claude to index a folder - indexing runs on your machine and a large drive takes a while the first time - and search it. Details and the full tool list are on the Reelary for AI agents page.
The search covers what the agent cannot cheaply do itself: what is visible in every two-second slice of every video, a written description of each shot, and with Pro what was said, who is on screen and where it was filmed.
What still happens in the cloud
The video does not move. The search results do: when the tool answers, Claude reads that text, and Claude runs on Anthropic's servers. For most libraries that is a list of file names and descriptions, much like pasting a folder listing into a chat. For client work under NDA it is worth knowing exactly what is in that list - see what an AI agent actually sees of your footage.
When you do not need any of this.
If you have a dozen clips, open them. If you want a summary of one recording, attaching it to a chat is simpler. And if you would rather not involve an agent at all, the Reelary app does the same search and cut with a search box and a timeline, and sends nothing anywhere.
Questions people ask about this
Can Claude watch the videos on my drive?
No. Claude does not read video files from your disk. With a local MCP server it can ask an index on your machine what is in them and get text back: paths, time ranges and descriptions. The index is built by a vision model running locally, not by Claude.
Does this work in Claude on the web?
Not with a local server. The claude.ai website cannot start programs on your computer. Claude Desktop and Claude Code can, and so can Cursor and Codex.
Why not just attach the videos to the chat?
A chat attachment is one file in one conversation, sampled into a few frames, and limited in size. A library is thousands of files that keep growing. Indexing has to happen once, where the files are, and then be queried many times.
What does it cost?
The Reelary app, the MCP server and the skills are free to install. Indexing and visual search are free and unlimited; MP4 export on the free plan is 720p with a watermark. Pro, a one-time purchase, adds dialogue search, people, places, NLE export and full-quality export.