TrimScript

What TrimScript does

TrimScript MCP server connects an MCP client to TrimScript’s local transcript-based video editor. An agent can read a transcript, identify repeated sections or misheard terms, and submit proposed changes. Suggestions appear in the open editor with a reason; they do not change the edit without user approval.

The workflow is intended for recorded speech. Selecting transcript words removes the corresponding media segment, while the editor shows the removed text as a restorable label. Built-in operations can remove detected “um” and “uh” fillers, shorten pauses longer than one second, and identify stray vocal sounds that were not represented in the transcript. Text corrections change captions and transcript content without changing the media.

TrimScript belongs in the Multimedia Process category for transcript-led work on talks, tutorials, podcasts, and interviews. It does not provide a multitrack timeline, screen recording, voice cloning, or collaboration features.

How it works

Transcription runs locally with whisper.cpp and currently supports English. The first run downloads a model of about 470 MB. TrimScript derives removed time from transcript items and places cuts between words or around pauses. Video exports use whole frames and audio is sliced to approximately millisecond precision, keeping sound aligned with picture across multiple edits.

The editor, command-line interface, and TrimScript MCP server read and write the same project JSON file next to the source video. The original media is not modified, and exports do not overwrite it. Users can review suggestions in the editor, listen across joins, restore removals, and adjust edit boundaries in 10 ms steps.

Setup and configuration

Install uv and ffmpeg on the machine that will process the media. The README gives platform-specific package examples: Homebrew on macOS, apt on Linux, and winget on Windows. Open a video with uvx trimscript open my-video.mp4.

To register the MCP server with Claude Code, run claude mcp add trimscript -- uvx trimscript mcp. The server operates locally with no account or upload required. If an external coding agent is connected, that agent and its model service can read the transcript.

The project is early-stage: end-to-end CI covers macOS, Linux, and Windows, but the editor itself has so far only been used on macOS. Development was done on Apple Silicon.

Tools and capabilities

TrimScript MCP server provides these tools:

• transcribe creates a local transcript.
• get_transcript reads transcript content.
• suggest_removal proposes a media removal.
• correct_text fixes recognized wording without altering media.
• list_edits lists current edits or suggestions.
• withdraw removes a proposed change.
• enhance_voice applies noise reduction, leveling, gentle compression, and a -14 LUFS target without shortening the recording.
• export produces MP4, WAV, SRT, VTT, or FCPXML output.

FCPXML keeps edit points adjustable in Final Cut Pro or DaVinci Resolve. The same editing actions are also available through the CLI, including filler removal, pause tightening, accepting suggestions, and export.

Limitations and notes

Transcription is English-only in the documented workflow, and recognizer output can mishear words or omit breaths and other sounds. The stray-sounds feature helps users inspect those cases, but decisions remain manual. Clean audio changes sound treatment rather than duration.

TrimScript is designed for single-track spoken-word recordings, not full timeline production. Use FCPXML when the project needs a conventional video editor. Agent edits remain suggestions until accepted, but undo in the editor only reverses changes made by the editor itself.