Refine video content using voice data with an editing agent
Place a raw video file in a folder and say "make a launch video from this," and a finished cut with unnecessary words and silence removed is returned. This tool performs editing using speech recognition results and screenshots of key moments, without directly viewing video frames. Because it is controlled by prompts, there is no need to touch menus or settings screens.
Editing based on voice data
The tool is designed so that the LLM does not directly analyze video frames. Instead, it uses a text file extracted via ElevenLabs Scribe, containing word-level timestamps and speaker information, as its primary input. It generates filmstrips and waveform graphs for specific time intervals only when visual verification is needed. This approach makes editing decisions using approximately 12KB of text and a few images, rather than processing millions of tokens of video data. This follows the same principle as browser agents using structured DOM instead of screenshots.
Automated editing tasks
It automatically removes filler words like 'um' and 'uh' from dialogue and silent intervals between takes. Color correction is applied to each segment, and a 30ms audio fade is added at every cut point to prevent pop noise. Subtitles are inserted by default in a two-word uppercase format, but the style can be freely changed. It uses HyperFrames, Remotion, Manim, and PIL to create animated overlays, with each animation handled by parallel sub-agents. The memory of every editing session is stored in the project.md file, allowing work to continue in the next session.
Installation and execution
Pasting the installation prompt into an agent with shell access, such as Claude Code, Codex, or Hermes, automatically handles cloning, dependency installation, ffmpeg integration, and ElevenLabs API key registration. For manual installation, clone the repository, create a symbolic link in the agent's skill directory, and install dependencies using uv sync or pip install -e .. ffmpeg is required, and yt-dlp can be optionally installed for downloading online sources. After entering the ElevenLabs API key in the .env file, run the agent from the folder containing the raw video.
Editing quality assurance process
Rendered output goes through a self-evaluation step before the user views it. It checks for visual jumps, audio pops, and hidden subtitles at all cut boundaries, and performs up to three rounds of correction and re-rendering if issues are found. A preview is provided only after this process is passed. The editing strategy is executed only after user approval, and it proceeds with editing after first checking, without making prior assumptions about the content type.
Pre-use considerations
This tool requires an ElevenLabs API key and operates in an environment where ffmpeg is installed. It requires an agent environment with shell access, such as Claude Code, Codex, Hermes, or Openclaw. The accuracy of video editing depends heavily on the quality of speech recognition, so raw videos with significant background noise may result in recognition errors. The README does not specify particular operating system or hardware requirements, but it is compatible with most environments where ffmpeg and Python dependencies can be installed. It is suitable for content requiring script-based linear editing rather than complex non-linear editing.
