talkthrough for OpenClaw

This skill wires the talkthrough MCP server into OpenClaw and teaches the workflow. Everything runs locally: recordings never leave the machine.

Setup (once)

Add the MCP server:

openclaw mcp add talkthrough --command uvx --arg --python --arg ">=3.11,<3.14" --arg "talkthrough-mcp[diarization,url]"

Requires uv (https://astral.sh/uv). First processing downloads a whisper model once (~460 MB for the default small).

Workflow

1. process_media(path) — idempotent by content hash; returns a compact summary with job_id (re-calls on the same file are instant). For one public video/audio URL use process_url(url): the source is downloaded once (the only network step), then everything below is local.
2. get_transcript(job_id) / search(job_id, "<word>") — orient; search covers speech AND on-screen OCR text.
3. get_moment(job_id, t0-2000, t1+2000) — evidence bundle per remark: transcript slice + up to 3 frames + OCR + wall-clock range.
4. extract_frame(job_id, at_ms, crop=...) — exact instant, native resolution, when keyframes miss the moment.
5. list_jobs() — recordings processed earlier remain queryable.

Timestamps: t_ms is video-relative; t_wall is real wall-clock time when the recording start is known — use it to correlate remarks with logs (±30 s window). Audio-only files (.m4a/.mp3/…) have transcripts but no frames; frame tools erroring on them is expected.

Full docs: https://github.com/korovin-aa97/talkthrough-mcp