Your agent can search your screen recordings

You record a clip of a bug. Two hours later you are in Claude Code, typing out a description of that same bug from memory, because the recording is a video file and your agent cannot watch video. The answer is sitting in your library. The agent just cannot reach it.

ClipCabinet fixes the reach.

What it is

ClipCabinet runs an MCP server. MCP (Model Context Protocol) is an open protocol that lets AI tools connect to outside data sources, and it is how Claude Code, Cursor, and a growing list of clients talk to things beyond your codebase.

You mint an API token, connect your client, and your agent gets tools for your recording library. It can list your clips, search across them, and read any single clip in full: the transcript of what you said, the summary, the frame captions describing what was on screen, and the clip's metadata. Screenshots work the same way, since they carry the same structured context as a recording.

The server runs over an HTTP transport that works out of the box with Claude Code and similar clients, so connecting does not involve local proxy gymnastics. One token, one config entry, done.

Recordings stop being attachments

Every other way of sharing a recording treats it as an attachment. You pick the clip, you paste the link or the file, you explain what is in it. The agent gets exactly what you handed over and nothing more.

Over MCP, the relationship flips. Your library becomes a data source the agent queries on its own. It does not need you to remember which clip showed the bug; searching for the clip is its job now. The screen recorders we compare against on our comparison pages do not offer this. Their recordings end as hosted videos behind a player, readable by humans only.

That flip is the whole point. A recording you have to describe is a memory aid. A recording your agent can query is context.

How it fits your workflow

Here is the actual loop.

First, connect. Open Settings, then API tokens. You can mint a token by hand, but the faster path is the connect cards: each supported client has a card that mints the right token and sets up the config for you. Claude Code, Cursor, VS Code, and Codex connect in one click. Settings labels each token with the client it belongs to, so you always know what is connected.

Then record like you normally would. Hit record in the extension, walk through the bug or the feature, hit stop. Processing starts on its own: transcript, summary, frame captions, embeddings. What that pipeline produces, and why it makes clips searchable at all, is covered in AI-searchable screen recordings.

Then ask. In your next agent session, type something like "find the clip where the websocket dropped right after login and tell me what the console said." The agent searches your library, opens the right clip, reads the transcript and the captions, and answers. You did not paste anything. You did not even switch windows.

The recording you made for a teammate this morning is the same recording your agent reads this afternoon. One capture, two audiences.

Why this makes you faster

Count what disappears.

You stop re-describing. The three paragraphs of "so when I hit submit, the form clears every field and nothing actually saves" become "look at my latest clip." The description was always a lossy copy of the recording anyway.

You stop re-watching. Without MCP, you are the bridge between the video and the agent: you scrub, you find the moment, you transcribe what you see into the chat. With MCP, the agent does the lookup while you keep working on something else.

And the context is durable. A pasted screenshot lives in one conversation and dies with it. A clip in your library is queryable from any session, today or a month from now. Record a bug once and every future debugging session can reach it.

Try it on one real bug

The setup takes a minute, and MCP comes with the free plan. Nothing about this loop sits behind a paywall.

Install the extension, record one real bug, and connect Claude Code from the connect card. Then start a session and ask about the bug without pasting a single thing. The first answer the agent fetches on its own is the whole pitch. The full tool reference lives at /docs/mcp. And if you want the wider survey of ways to hand work to an agent, with MCP ranked against the alternatives, read Best ways to share what you're working on with your AI agent.