zoom-editor

Turn a meeting recording and its transcript into short, captioned videos ready for social media: a single clip, a trailer of about 30 seconds, or a longform best-of. Each comes as a plain cut and, optionally, as 1:1, 9:16 and 16:9 versions with burned-in captions, a waveform, spectrum bars and a headline.

It runs as one service on port 8093, serving an MCP server (streamable HTTP at /mcp) for AI agents and a REST API (/api/v1) for scripts. It encodes on the GPU with NVIDIA's encoder (h264_nvenc) when one is available and falls back to the CPU (libx264) otherwise.

Zoom is the motivating case because Zoom cloud recordings come with an audio_transcript_*.vtt. Any video with a WebVTT or SRT transcript works.

How an agent uses it

1. create_project: point it at the video (URL, or a path under the media root) and the transcript.

2. get_transcript: read the cues and decide which moments are worth sharing. The service doesn't pick highlights itself; the calling agent (or person) does, so no LLM key is needed here.

3. Render:

   • render_clip: one moment.

   • render_trailer: one whole sentence from each moment, the one that best matches its hook, played in recording order.

   • render_longform: every moment in full, with the headline changing for each.

   • render_social: no cut; the social sizes of the whole video, for a clip you already have. caption_start says where it sits on the transcript's timeline, and headline_timeline can change the headline over time.

4. get_job until the job's status is done, then download the file URLs.

Times can be given in seconds or as HH:MM:SS.mmm, measured on the source video's timeline. Captions never cut a sentence mid-way, and trailer snippets always start and end on sentence boundaries.

Related MCP server: atsurae

Run it

docker pull ghcr.io/isleprince/zoom-editor   # or build: docker compose up -d --build
docker compose up -d                          # GPU passthrough requested; works without one too
curl http://localhost:8093/api/v1/health

/api/v1/health reports which encoder was chosen and why. The choice comes from a short test encode at startup, not from checking nvidia-smi. For NVENC inside Docker, the NVIDIA container runtime must pass the video capability, which the compose file requests.

Connect an MCP client to http://<host>:8093/mcp (streamable HTTP).

REST

GET    /api/v1/health
POST   /api/v1/projects                 {video_url|video_path, transcript_url|transcript_path|transcript_text, name}
POST   /api/v1/projects/upload          multipart: video, transcript?, name?
GET    /api/v1/projects[/{id}]          DELETE /api/v1/projects/{id}
GET    /api/v1/projects/{id}/transcript ?start=&end=
POST   /api/v1/jobs                     {project_id, type: clip|trailer|longform|social, moments:[{start,end,label,hook?}],
                                         sizes:["1x1","9x16","16x9"], subtitles, headline, accent, bg, quality,
                                         caption_start, headline_timeline:[{t0,t1,text}]}   (last two: social)
                                        (a clip may instead give start/end/headline at the top level)
GET    /api/v1/jobs[/{id}]              GET /api/v1/jobs/{id}/files/{name}

Develop

pip install -e '.[test]'
pytest -q        # generates its own test video; needs an ffmpeg that has drawtext

The layout and transcript rules were ported from the nokemo.com Zoom-to-social pipeline (render_social.py, render_compilations.py) and behave the same.