chatmonteur

ChatMonteur is an extensible, agent-orchestrated editing studio for real talking-head recordings such as vlogs, tutorials, and explainers. It does not generate video from scratch; it edits existing raw footage through a pipeline that orchestrates established open-source tools: auto-editor, faster-whisper, HyperFrames, and ffmpeg. The system is driven by a coding agent (Claude Code, Codex, or similar). The user drops in raw footage and describes the desired edit in plain language. The agent then runs the pipeline, which transcribes the audio, cuts pauses and stumbles by meaning, generates subtitles, adds motion graphics, sound, and color grading, and renders the final file. Two quality gates are enforced throughout the workflow. The plan gate scores the visual plan before any frames are rendered and rejects outputs that would read as mechanically cut — for example, stretches over 90 seconds with no visual event, text on screen more than 60 percent of the runtime, repeated captions, or three identical zooms in a row. The file gate re-opens every encoded artifact at sampled timestamps (10, 35, 65, and 90 percent), checks audio levels for silence and clipping, and compares runtime against what the encoder was given. Any broken file stops delivery and produces a sibling `.qc.json` report. The editorial decision knowledge lives in the `skills/` directory, written as an auditable knowledge base rather than embedded prompts. This includes guidance on where to cut, how to pad cuts, when to zoom, and how to balance the music bed. All engineering thresholds are documented in `engineering-facts.md` with explanations of why each rule holds. A typical workflow produces a mechanically valid draft with one command, after which the user talks to the agent to finish the edit — cutting filler, adding captions and graphics, and refining sound. The agent presents the required approval gates and promotes only completed edits to a master file. The project is free and local-first. `faster-whisper` runs locally without paid API keys. Additional transcription backends can be added as plugins. Cross-platform encoding is supported with NVENC fallback to libx64, VideoToolbox, or QSV. Status is v0.1. The mechanical pipeline works end-to-end, and the agent-driven workflow has been tested on a full YouTube-channel edit from raw footage to a checked master.