PaulBratslavsky

Roughcut AI Local First Editor — AI skill for Claude Code

AI community

Open-source, local-first AI video editor: on-device transcription (whisper.cpp), local LLM rough cuts (Gemma via Ollama), transcript-based editing, NLE export, and an MCP tool surface Claude can drive.

How to install Roughcut AI Local First Editor

This entry records only its repository, not the path inside it, so there is no exact command to give. Open PaulBratslavsky/roughcut-ai-local-first-editor and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Roughcut AI Local First Editor does

Alternatives in AI

Repomix — 📦 Repomix is a powerful tool that packs your entire repository into a single, AI-friendly file 22.7k ★
Inference Serving — ai-research-skills vLLM, SGLang, TensorRT-LLM, llama.cpp 5.4k ★
Surface — Show ranked attack surface for a target from its recon manifest + hunt memory 3.8k ★

README

RoughCut

**An open-source, local-first, AI-powered video editor.** Raw talking-head footage goes in; a clean rough cut comes out — **entirely on your machine**. Transcribe on-device, let a local LLM remove silences, filler words, and bad takes, refine by editing the transcript or chatting with your footage, then export to Premiere, Final Cut, or DaVinci Resolve. No upload, no subscription, no account.

Uniquely, the editor exposes its tools over [MCP](https://modelcontextprotocol.io), so Claude Desktop or Claude Code can drive it — the same tool registry powers the local agent loop and the UI. One tool set, three callers.

Mission: prove that a paid cloud AI service can be replaced by owned, local-first software. See [`docs/01-product.md`](docs/01-product.md).

![RoughCut — terminal-style editor with transcript, preview, and timeline](docs/screenshot-dark.png)

Highlights

Edit text, not timelines — strike a sentence in the transcript and the cut follows; click a word to seek; ←/→ steps word-by-word with an audio cue
AI rough cut in one click — on-device whisper transcription + silence/filler/take detection; composes with your manual edits and is one undo step
"Make this 20 minutes" — semantic duration planner ranks segments by importance (local embeddings) and batches the cut, via chat or one tool call
Hybrid transcript search — BM25 + local embeddings; "find the part about…" works on meaning, not keywords
NLE hand-off — Premiere/Resolve XML, FCPXML, EDL, OTIO, SRT, or a rendered MP4
Claude can drive it — 30+ MCP tools with batch editing; every external edit is tagged and undoable
Zero egress — the only network use is user-triggered model downloads; everything is verifiable with a network monitor