MCP Server for Bulk Video Editing from Claude Code / Claude Desktop

Claude handles the mechanics — taste stays with you

The VideoBuff MCP server is a tool for driving timeline operations in bulk from Claude Desktop or Claude Code (CLI). The intended use case is simple: finishing "I know exactly what I want, but doing it by hand is tedious" tasks in a single prompt. Typical examples:

• Place 30 imported images on the timeline at 3 seconds each with 0.5-second crossfades between them
• Unify the font size of every text clip on the timeline to 48
• Strip every silent region (< -40 dB, ≥ 0.5 s) from the audio track
• Overlay "#01", "#02", … numbering text on each clip, bottom-left
• Apply the same color parameters across every clip

It is NOT designed for vague creative prompts like "make this look cool" or "edit this into a snappy short." Claude only reads timeline metadata (clip list, durations, text content, color parameters) and issues mechanical operations — it cannot "see" the preview and judge whether the result looks good.

The sweet spot is a division of labor where you decide the taste (color, pacing, composition) and Claude executes the mechanical follow-through. Think of it as "Premiere Pro or Final Cut macros you can write in natural language" rather than AI creativity for video.

The MCP server operates VideoBuff through a browser tab, so closing the tab immediately disconnects the integration — Claude is never granted persistent access. Manual editing continues to work normally; your mouse and keyboard are not locked out while Claude is active.

Installation: Claude Desktop

If you use Claude Desktop, a one-click installable .mcpb bundle is available from the VideoBuff site footer or the settings dialog. Double-click the downloaded .mcpb file to open Claude Desktop, which will prompt you to confirm adding the extension.

Once approved, VideoBuff-specific tools (add clip, edit text, open export dialog, and so on) become visible to Claude. Open the VideoBuff editor tab and send Claude a prompt — it will read the tab state and call the relevant tools.

When a new version ships, download the latest .mcpb and install it the same way; it overwrites the previous install.

Installation: using it from Claude Code (CLI)

For Claude Code users, the VideoBuff MCP server slots into your existing CLI workflow as a stdio server. Expand the .mcpb bundle, then register dist/index.js as a stdio command in your Claude Code MCP config (e.g. ~/.config/claude-code/mcp.json) and the integration is active.

With claude code running in a terminal and the VideoBuff tab open in your browser, a prompt like "line up every video in this folder on the timeline, 5 seconds each" drives the edit directly from the terminal. This is the closest fit for developers who want to slot video editing into the same CLI tier as git, ffmpeg, or make.

If multiple Claude Code sessions try to operate the same VideoBuff tab, a leader election lets exactly one session become active while others follow along, sharing the same edit state. You can seamlessly take over from a crashed terminal session without losing progress.

See the README inside the distributed bundle for example config snippets.

Prompt examples: knocking out repetitive work in one shot

The sweet spot for VideoBuff MCP is "monotonous but high-volume" work that should finish in a single prompt. Both Japanese and English prompts work.

Bulk layout

• "Line up every imported image on the timeline at 3 seconds each with 0.5-second crossfades"
• "Overlay '#01', '#02', … on the 20 clips in the video track, bottom-left, size 32, white"
• "Unify every text clip to font size 48, color #ffffff, with a drop shadow"

Bulk processing

• "Strip every silent region (< -40 dB, ≥ 0.5 s) from the audio track"
• "Bump saturation by +10 and contrast by +5 across every clip"
• "Set every even-index clip to 1.5x speed"

Selective operations

• "Delete every clip shorter than 3 seconds, then ripple-close the gaps"
• "Remove every text clip whose content contains 'DRAFT'"
• "Move every clip that starts after 30 s into track 2"

Claude reads the current timeline state (clip list, durations, text content, color parameters) via tools before acting, so state-dependent prompts like "make every existing text 10 pt larger" work correctly.

What tends to fail is abstract prompts: "make it feel cool," "TikTok-style," "cinematic vibe." Claude cannot translate those into concrete operations, so you get either nothing or a miss. Translate the direction yourself into concrete operations ("set every clip to 1.5 s and punch-in between them") before handing it over.

Exports stay manual. Resolution, codec, and save location all need a human choice, so Claude opens the export dialog and hands off: "pick resolution and destination, then click Export."

What it can't do — calibrating expectations

Calibrating expectations matters — misuse the tool and you'll end up disappointed. The constraints at this point:

No visual or audio judgment. Claude cannot "see" the preview to evaluate layout or "listen" to waveforms to align narration. It only touches metadata (clip list, durations, text content, color parameters, measured font sizes) and executes mechanical operations. Whether the result "looks good" is something only you can verify by watching the final preview.

Abstract / aesthetic prompts are weak. "Make it cool," "professional," "TikTok-style," "cinematic color grade" cannot be translated into concrete operations, so they either no-op or miss. Concrete parameters like "saturation +15, contrast +10, color temperature -200" work reliably.

File I/O requires a human click. Both import and export flows route through the browser's native file picker / save-location dialog. Claude opens the dialog and hands off to you. Direct filesystem access is intentionally not granted to Claude.

Chain complex edits step by step. Packing "cut → text → color → export" into a single prompt is less reliable than stepping through them one at a time. Large batch operations on the order of tens of clips are fine; chaining five different kinds of operations into one final form is where context can drift.

Closing the tab disconnects the integration. The VideoBuff tab must be active for Claude to operate. This is not the right tool for leaving long, fully-autonomous edits running in the background unattended.

Privacy: what is actually sent

Video, audio, image, and project data are never sent to any server, even while Claude integration is active. All editing happens inside your browser, and Claude only reads the browser state through a local MCP server.

Only anonymous usage stats may be sent for improvement. These are limited to tool names, success/error outcomes, coarse duration buckets, the MCP version, and your OS family. File contents, project names, and text you have entered are never included. To understand usage trends, a single anonymous UUID is stored on your device and attached as an identifier — with no personally identifiable information.

You can toggle the stats off anytime from Settings — the gear icon at the top right of the editor. See the privacy policy for full details.