The talkthrough MCP server turns a local video/audio file into queryable\nstructured data: timestamped transcript segments, scene keyframes, OCR'd\non-screen text, and wall-clock anchoring. No LLM inside — you bring the\nreasoning; it brings the evidence. Everything is lazy and token-budgeted:\nnever ask for more than the moment you are analyzing.
\nThe talkthrough MCP server must be connected (tools like\nprocess_media / get_transcript are visible). If not, tell the user to\ninstall it: claude mcp add -s user talkthrough -- uvx talkthrough-mcp\n(see the repository README for other clients).
process_media(path) — idempotent by content hash;\nre-calls on the same file return instantly. Long videos take minutes and\nstream progress. The summary gives you job_id, counts, wall-clock, and\na transcript preview — do NOT dump anything else eagerly. Multi-person\nrecording (meeting/interview)? Add diarize=true — even when the ask is\njust \"summarize\", speaker structure is part of meeting analysis — and — whenever the\nheadcount is known — num_speakers=N (the main accuracy lever): segments\nget S1/S2/… labels and the summary a talk-time roster. On an\nalready-processed job the amend re-runs ONLY diarization (no\nre-transcription) — still minutes on long recordings.get_transcript(job_id) (paginate via next_start_ms when\ntruncated) or search(job_id, \"<distinctive word>\") to jump straight\nto the relevant moments (searches speech AND on-screen OCR text). Multi-word\nsearch defaults to match_mode=\"all_words\"; use \"any_word\" for broader\nlexical recall.get_moment(job_id, t0-2000, t1+2000) — one\ncall returns the transcript slice + up to 3 unique frames + their OCR\ntext + the wall-clock range. This is the workhorse; describe observed\nfrom the returned pixels, never from imagination.get_frames(at_ms=...) for nearby keyframes;\nextract_frame(job_id, at_ms, crop={x,y,w,h}) for an exact instant at\nnative resolution (keyframes capture scene changes + a 1 fps floor, so\nsub-second moments can fall between them).label_speakers(job_id, labels={\"S1\":\"Name\"}, evidence={\"S1\":\"intro or frame proof\"}). Saved names appear in later\ntranscript, moment, and search calls while raw S1/S2 labels remain.list_jobs() — the store persists; a file\nprocessed yesterday (even via CLI) is queryable by job_id today.Every timestamped result carries t_ms (video-relative) and, when the\nrecording start is known, t_wall (ISO 8601 real time). Copy t_wall\nVERBATIM from the payload — never compute it from t_ms yourself\n(hand-derived wall-clocks drift by whole hours). Use t_wall to\ncorrelate remarks with server/app logs (±30 s grep window). If\nwall_clock is null or low-confidence, ask the user when the recording\nstarted and re-anchor: process_media(path, recorded_at=\"<ISO 8601>\", force=true).
Prefer the server prompts when the task matches — they encode the full\nmethod: bug (one recording → evidence-backed GitHub issue draft; silent,\nnarration-free recordings welcome), triage-recording (screencast →\nfindings JSON per the contract in examples/output-contract.schema.json),\nspec-from-workshop, backlog-from-demo, meeting-actions (audio-only\nfriendly), correlate-with-logs.
S1/S2, ordered by first voice). Mapping\nthem to names is YOUR job: self-introductions, vocatives, the attendees\nlist — and on video jobs the screen check is MANDATORY: for every label\nyou map, get_frames(at_ms=<that label's longest_turn_at_ms from the roster>) and read the meeting-app name plates, the recording's title\ncard, the active-speaker highlight BEFORE asserting the mapping. STT\nhomophones lie about name spellings (spoken \"profit\" vs on-screen\n\"Prophet\") — trust OCR/frames over the transcript for names. State the\nmapping explicitly and mark unmapped labels \"unidentified\".\nRoster name_candidates are raw OCR hints, not identities: they may be\nUI text, a job title, or somebody else's name. Inspect the cited frame and\npersist only defensible mappings with label_speakers; never auto-save a\ncandidate.\ndiarize=true needs the [diarization] extra — its absence produces an\nactionable install-hint error.t_ms (+ t_wall\nwhen known) + the frame files you actually inspected.language + language_probability). Garbled transcript or low/wrong\ndetection → re-call process_media(path, model=\"large-v3-turbo\", force=true) (best multilingual quality) or pin language=\"…\"; domain\njargon → pass vocabulary=\"Term1, Term2\".