talkthrough-mcp

Quickstart · Tools · Benchmarks · FAQ · Troubleshooting · Changelog · Contributing

Don't write a bug report. Record it.

Give Claude Code or Codex a narrated .mov/.mp4 — or a public video link — and talkthrough turns it into searchable transcript, exact frames, OCR and wall-clock timestamps, locally — so your agent writes an evidence-backed issue draft or investigates the fix.

Also works for meetings, workshops, product demos, and production incidents.

Illustration of a /talkthrough:bug run: a recorded checkout bug is indexed locally, the evidence is found, and a ready-to-file issue draft is assembled

Illustration — an animation of the /talkthrough:bug storyline, not a screen capture: the recording is indexed locally (transcript · keyframes · OCR · wall-clock), the evidence checkpoint is assembled, and the draft is ready to file with your own tracker tooling (gh, Jira, a GitHub MCP server — talkthrough itself never leaves your machine and never posts anything).

Real runs, no animation:

• ▶ Watch the demo with sound (1:18) — an unedited session: a narrated recording goes in, a ready-to-file bug-report.md comes out.
• Silent recording → issue draft — the whole thing as files you can re-run: a Playwright-recorded, audio-free .mp4 and the unedited agent output. Every number is reproducible — processing that file gives job_id 8703a66bbe77a7d0 (the job id is the sha256 prefix, so shasum -a 256 predicts it), 17 keyframes / 4 unique, and 0 transcript segments because there is no audio track.

Quickstart

One command, no system dependencies: ffmpeg falls back to a bundled build, OCR is pip-only, and whisper models download themselves on first use. The only prerequisite is uv (brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh).

Cold setup has two separate stages: uvx first resolves a compatible Python and the pinned server environment, then the first process_media downloads any missing media/model assets. A plugin update can create a new environment and, without system ffmpeg, fetch the ~80 MB bundled ffmpeg again; shared Whisper/OCR/diarization caches and warm, network-free jobs remain reusable. See Troubleshooting.

Claude Code

Two install paths — pick one, not both (the plugin already includes the server; installing both would register it twice):

Server only — the 9 tools + 6 prompts, and nothing else on your system. Choose this for a minimal setup, or when you manage MCP servers yourself across several clients:

claude mcp add -s user talkthrough -- uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]"

Full plugin — the same server, plus native slash commands (/talkthrough:bug, /talkthrough:triage-recording, …) that handle the ceremony for you, a ready-made triage subagent, and an agent skill that teaches Claude the workflow. Choose this for the best out-of-the-box experience:

/plugin marketplace add korovin-aa97/talkthrough-mcp
/plugin install talkthrough@talkthrough

Every other MCP client

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

[mcp_servers.talkthrough]
command = "uvx"
args = ["--python", ">=3.11,<3.14", "talkthrough-mcp[diarization,url]"]

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

{
  "mcp": {
    "servers": {
      "talkthrough": {
        "command": "uvx",
        "args": [
          "--python",
          ">=3.11,<3.14",
          "talkthrough-mcp[diarization,url]"
        ]
      }
    }
  }
}

{
  "mcp": {
    "talkthrough": {
      "type": "local",
      "command": [
        "uvx",
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ],
      "enabled": true
    }
  }
}

extensions:
  talkthrough:
    enabled: true
    type: stdio
    cmd: uvx
    args: ["--python", ">=3.11,<3.14", "talkthrough-mcp[diarization,url]"]

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

{
  "mcpServers": {
    "talkthrough": {
      "command": "uvx",
      "args": [
        "--python",
        ">=3.11,<3.14",
        "talkthrough-mcp[diarization,url]"
      ]
    }
  }
}

{
  "context_servers": {
    "talkthrough": {
      "source": "custom",
      "command": {
        "path": "uvx",
        "args": [
          "--python",
          ">=3.11,<3.14",
          "talkthrough-mcp[diarization,url]"
        ]
      }
    }
  }
}

Any other MCP stdio client uses the same server command: uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]". Per-engine folders with exactly these snippets plus verification steps live in integrations/; agents can self-install via llms-install.md.

Who said what (speaker diarization) — included in the configs above

Multi-person recordings (meetings, interviews, panels) can carry S1/S2/… speaker labels. Every install button, snippet, and the plugin above already ship the [diarization] engine, so asking your agent "who said what" just works — diarization itself still runs only when requested per call (process_media(path=..., diarize=true, num_speakers=<count if known>)), and its models download once on first use.

Prefer the minimal server without the diarization engine? Use uvx --python ">=3.11,<3.14" talkthrough-mcp as the command instead (the MCP registry entry also resolves to this lean form) — an explicit diarize=true will then answer with the one-line install fix. Details in Speakers.

Upgrading from 0.3.x

Regenerated configs and the plugin carry [diarization,url]. A config you wrote by hand for 0.3.x — uvx --python ">=3.11,<3.14" talkthrough-mcp, or a pin without the url extra — upgrades the server in place and lists the new process_url tool, but only direct https:// media links work until the extra is there: YouTube and video pages answer with the one-line install fix. Add url to your command (uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]"), restart the client, and check with talkthrough-mcp --version (0.4.1+), which names the extras the environment has; the server logs the same line to stderr at every start, so your client's MCP log shows it too.

Local checkout (development)

git clone https://github.com/korovin-aa97/talkthrough-mcp
claude mcp add talkthrough -- uv run --directory /path/to/talkthrough-mcp talkthrough-mcp

Then, in your agent:

Process ~/Desktop/recording.mov and triage it — or just invoke the triage-recording server prompt.

Tools

Every tool description ships 10+ usage examples, so agents pick the right tool without extra prompting.

Server prompts (slash commands in MCP clients)

The same prompts live as plain files in examples/prompts/ if your client doesn't surface MCP prompts. The findings contract used by triage-recording is examples/output-contract.schema.json.

Works as a skill too (no MCP required)

The same workflow ships as a cross-engine Agent Skill at .agents/skills/talkthrough/ — Claude Code, Codex CLI ($talkthrough), Cursor, Copilot, Gemini CLI, Goose and other SKILL.md-compatible tools read it. Agents without MCP wiring can drive the CLI directly: talkthrough-mcp process recording.mov --json prints the same summary the MCP tool returns, and the job store is shared either way.

Wall-clock anchoring

Every timestamped result carries both t_ms (video-relative) and t_wall (ISO 8601 real time) once the recording start is known. Resolution ladder:

1. recorded_at parameter (agent/user override) → confidence exact
2. QuickTime com.apple.quicktime.creationdate tag, carries the local timezone (QuickTime Player recordings; ⌘⇧5 wrote it before macOS 26) → high
3. Container creation_time tag (UTC) → medium — macOS 26+ ⌘⇧5/ReplayKit screen recordings land here (no creationdate tag anymore); pass recorded_at= when local-tz t_wall matters
4. File mtime minus duration (recorders finalize files at recording END) → low
5. Nothing → tools still work with relative t_ms only

Why it matters: "the upload spinner froze here" becomes a ±30 s grep window in your server logs.

Speakers (optional diarization)

With the [diarization] extra installed (included in every generated config — see Quickstart), process_media(diarize=true) labels who said what — locally, like everything else here (sherpa-onnx runtime, no torch, no accounts, no GPU):

• Speakers become S1, S2, … in order of first appearance; new diarized jobs split speaker changes at word boundaries, while old jobs continue to report honest segment-level precision. Every transcript segment gets a speaker, and the tools surface it everywhere — roster with talk time in get_transcript, speakers_in_range in get_moment, speaker on search hits, S1: prefixes in the text/SRT formats, a speaker count in list_jobs.
• Know the headcount? Pass num_speakers. Clustering toward an exact k removes the main failure mode of unknown-count mode (similar voices merging or one voice splitting). It is a target, not a guarantee: the clusterer can converge on fewer clusters than k, and a re-run that changed nothing says so in the payload (labels_changed: false). Agents are instructed to pass the headcount via the tool guidance; do the same in your own calls.
• Already processed a recording? Calling process_media(diarize=true) on it re-runs only diarization — whisper is not re-run, and labels land in the existing job. Same for changing num_speakers. The diarization stage itself still re-scans the full audio: minutes on long recordings.
• Full rebuilds keep named jobs safe. If a job has active or pending names, force=true also requires diarize=true; otherwise the call refuses before changing stored data. A successful force rebuilds in staging and moves every previous identity to pending review against the fresh roster. Any processing or commit failure leaves the prior manifest and frames intact.
• Labels start anonymous. After checking self-introductions, vocatives, or video evidence, call label_speakers to preserve a verified mapping such as S1 → "Alice" across sessions. The roster can expose bounded OCR name_candidates, but those are raw hints and are never saved automatically. The raw label remains present beside speaker_name. If a later diarization amend changes the labels, verified names stop being active and move to bounded speaker_names_pending_review evidence instead of being silently lost. Re-check the current roster and explicitly confirm or remove each affected label with label_speakers.
• Video jobs produced before 0.3.1 keep their original flat OCR and can return name_candidates_note to explain why hints are absent. They remain fully readable without migration; force=true, diarize=true regenerates line-aware OCR while preserving saved identities for review.

Models download once (~47 MB total) from pinned, checksum-verified URLs into ~/.talkthrough/models/; warm runs are zero-network like the rest of the pipeline. Speed on an M-series CPU (4 threads): a 26-minute meeting diarizes in about 2 minutes (RTF ≈ 0.08), on top of the transcription time. Memory: expect on the order of 1–1.5 GB peak RSS while an hour-plus meeting is being diarized (measured on a real 73-minute recording); it is released when the stage completes.

The default won a real-meeting accept-eval (RU/EN/ES + a 3-speaker 26-minute meeting): it was the only candidate to isolate all three real voices at num_speakers=3, at 2× the speed of the runner-up.

Pick an alternate embedding model (or point at your own .onnx file for offline machines) via TALKTHROUGH_DIARIZATION_EMB_MODEL; tune the unknown-count sensitivity via TALKTHROUGH_DIARIZATION_THRESHOLD (see docs/TROUBLESHOOTING.md). Honest quality notes live in Limitations.

Privacy

Everything runs locally: your recordings never leave your machine, speech is transcribed by a local whisper model, OCR and speaker diarization are local ONNX inference, and there is no telemetry. For local files the only network access is one-time tool/model downloads (ffmpeg build, whisper model, OCR models, diarization models — the latter pinned by URL + sha256). The one deliberate exception is process_url: it downloads the public source you name from its provider or CDN, once, and nothing else — no media ever goes up, no cloud STT or LLM is called, and after that download every tool on the job is network-free again. The raw URL (which may carry signed tokens) is not stored: the job keeps a hash, the public provider id or host and a bounded title. Diarization keeps no voiceprint database: voice embeddings live only in process memory, and only anonymous turn labels (S1/S2) land on disk. Your agent sees only the payloads the MCP tools return (text and selected frames) in your existing session; talkthrough itself makes no LLM calls.

Languages

Narration in any of Whisper's ~99 languages works: the language is auto-detected per recording, and the summary reports both language and language_probability so agents can tell a confident detection from a shaky one (silence or music at the start can fool the detector — pin it with language="ru" and force=true when that happens). Speaker diarization is acoustic — it fingerprints voices, not words — so it is language-independent and works across all of those languages unchanged.

Pick the model for your languages — per call (model= parameter, agents do this themselves when a transcript comes back garbled) or as the server default (TALKTHROUGH_WHISPER_MODEL):

Tips that work in every language: pass product names via vocabulary="Term1, Term2" (biases the decoder so jargon survives), and note that the workflow prompts instruct agents to write digests in the narrator's language while keeping quotes verbatim — the server never translates (exact quotes are evidence; translation is the agent's job).

On-screen text (OCR) defaults to RapidOCR's Latin + Chinese models. For other scripts set TALKTHROUGH_OCR_LANG to your language — ru/uk (→ the eslav pack), ja, ko, ar, hi, el, th, or any RapidOCR pack name like cyrillic — and reprocess with force=true; the matching recognition model downloads once. Spoken-language support is unaffected either way.

Configuration

CLI

The pipeline is also a CLI — useful for pre-processing long recordings outside an agent session (the store is content-addressed, so the agent then queries the same job instantly):

talkthrough-mcp process ~/Videos/long-session.mov   # prints the summary
talkthrough-mcp process demo.mov --json             # machine-readable
talkthrough-mcp process sync.m4a --diarize --num-speakers 3   # who said what
talkthrough-mcp process-url "https://youtu.be/nHfGfEiVdE8"    # one public URL, downloaded once
talkthrough-mcp gc --keep-days 30                   # clean the job store (sources go with their jobs)
talkthrough-mcp serve                               # stdio MCP server (default)
talkthrough-mcp --version                           # package version + which extras this environment has

--json keeps stdout machine-readable on failure too (0.4.1), including a missing argument or unknown option: the process exits with code 2, stderr carries the human error: … line, and stdout carries one JSON document, {"error": {"type": "UnsupportedUrlError", "message": "…"}}. --version also says which optional extras the environment has — the quickest check when a hand-written config launches the minimal server (a launcher without the url extra) that advertises process_url but can only read direct media links; the server logs the same line to stderr at every start.

First run notes: missing system ffmpeg triggers a one-time static-ffmpeg download; the first transcription downloads the whisper model (~460 MB for small); both are cached. After that, expect roughly 3× faster than real time on an Apple-Silicon CPU with the default model, OCR included (a 2-minute clip processes in ~40 s) — and instant re-runs on the same file. Progress streams as MCP progress notifications, and the CLI prints stage lines. More: docs/TROUBLESHOOTING.md.

Windows

CI runs lint, the unit suite, a full CLI smoke, and a diarize smoke on windows-latest (static-ffmpeg Windows build, whisper tiny transcription, OCR, the instant idempotent re-run, and a speaker-roster assert through the native sherpa-onnx stack). Notes: the per-job lock always serializes threads; POSIX also uses fcntl for cross-process locking. Quote paths with spaces (uv run talkthrough-mcp process "C:\Videos\Screen Recording.mp4"). If something breaks, please open an issue.

Supported inputs

Video: .mov .mp4 .webm .mkv .ogv — audio-only: .m4a .mp3 .wav .ogg .flac (transcript tools only; frame tools explain why they're unavailable).

URLs (via process_url, since 0.4.0): a direct https:// link to one of those media files; one public YouTube video (watch, youtu.be, shorts, a completed live); or any public video page yt-dlp can read — public Instagram reels, TikTok, Wikimedia Commons file pages (each verified on release day), the rest of yt-dlp's ~1800 site extractors and pages with a plain HTML5/HLS player (the [url] extra, which the generated configs above already carry, brings yt-dlp). Always anonymous: a site that demands a sign-in from anonymous clients (Vimeo does, with this yt-dlp) is refused with the reason. The source is downloaded once, kept inside the job, and never re-fetched for later questions. Not supported: playlists, channels, active live streams, private, members-only, age-restricted or DRM-protected videos, cookies or logins — a site that hides a video behind a login or a bot wall answers with a clear refusal, not a workaround (Instagram in particular rate-limits anonymous access). You are responsible for having the right to download and process what you point it at; talkthrough does not bypass any restriction.

Limitations

Honest edges, so you can decide fast:

• Speaker labels are word-level and opt-in on new jobs. Each Whisper word is assigned by maximum overlap with the diarization turns, so fast exchanges split without losing the raw label. Old jobs remain readable and honestly report attribution_precision="segment"; reprocess with force=true to add word timings. Sub-second interjections ("yeah", "mhm") can still be absorbed when the diarization engine does not detect a separate turn, and heavy crosstalk degrades clustering (the segmentation model tracks at most 2 simultaneous voices). Quality is pyannote-3.x-generation. The comfort zone without hints is roughly 2–8 speakers; pass num_speakers whenever the headcount is known — it removes the worst failure mode at any size, and it is the way to go for large meetings (10+).
• URL ingestion covers one public video at a time, without logins. Direct HTTPS media links, single public YouTube videos and public video pages yt-dlp can read; playlists, channels, live streams, gated or DRM content and anything behind a login or a bot wall are refused with a reason. Sites change; a page that worked yesterday can need a newer yt-dlp tomorrow. A provider's upload date is not a recording time, so URL jobs have wall_clock: null unless you pass recorded_at.
• Memory: budget about 2 GB for a cold run. Whisper, the OCR models and the frame pass live in one process. A 78-second video on tiny with OCR peaked at 1.6 GB RSS during the 0.4.0 release QA (download included); the default small model needs more, larger models proportionally so. An 8 GB laptop copes; on anything tighter keep the model small or run the CLI ahead of the agent session.
• Keyframes + transcript, not motion analysis. A glitch between scene changes can be invisible in the frame set; extract_frame re-checks any instant, but frame-by-frame motion reasoning is your multimodal model's job.
• STT quality tracks the model you pick. The default small favors speed; non-English narration wants model="large-v3-turbo" (see Languages).
• OCR reads crisp UI text well; tiny or low-contrast print is best-effort.
• Wall-clock confidence depends on recorder metadata — worst case pass recorded_at= (see the ladder above).
• Windows caveats — POSIX lock degrades to a no-op; see the Windows section above.

How it compares

FAQ

Why not just upload the video to a multimodal model (e.g. Gemini)? For a short, non-sensitive clip — do that. The trade-offs appear with length and sensitivity: an hour of screen recording costs on the order of a million tokens per question, the file leaves your machine, and you still can't map a remark to 14:32:07 UTC to grep your server logs. talkthrough indexes once, locally, then answers any number of follow-ups from the index.

Why not screenpipe? Different job. screenpipe is an always-on recorder of your machine going forward (commercial license). It can't open the .mov a teammate or customer just sent you. talkthrough analyzes any file it's handed — the two compose fine.

There are agent skills that "watch" videos. Why a server with an index? Watch-style skills push a budgeted frame dump into the context window (and go sparse on long videos), often call cloud STT for the audio, and keep nothing. talkthrough builds a persistent local index — transcript + OCR, full-text searchable — retrieves exact frames lazily, anchors everything to wall-clock time, and answers the next question without reprocessing.

I use Jam for bug reports — do I need this? Keep Jam for browser bugs: console+network captured at record time is great evidence. talkthrough covers what a browser extension can't — desktop apps, mobile screencasts, ops incidents, meetings, any file — with no account, and correlates with server-side logs via wall-clock time.

Which agent model do I need to drive this? For v0.4.0 the six model configs below ran 61 isolated behaviour cells on URL ingestion (YouTube, a TikTok page, a speechless Instagram reel, playlist refusal, the missing wall clock) and the 0.3.2 integrity fixes; every cell passes on every runner after two product fixes the first attempts exposed (Codex needs a per-tool approval for the open-world process_url, and the server now says itself when a URL job has no wall clock). For v0.3.0 we ran 210 isolated agent cells across 6 model configs (Claude haiku/sonnet/opus, Codex gpt-5.5 at two reasoning efforts, and gpt-5.4-mini) and 35 logical scenarios on 5 real recordings plus safety and speaker-label fixtures. All 102 LLM-judged full-grid results and every mechanical zero were manually audited; the 30 new speaker behavior runs passed, and old-server control left 0 release-caused regressions. This is a model-drift snapshot, not a leaderboard: see the current matrix in docs/MODEL-NOTES.md. The chart and narrative in benchmarks/ remain the historical v0.2.0 snapshot.

Can't I just script ffmpeg + whisper myself? Yes — that's exactly this pipeline. What you'd be rebuilding: scene-change detection with perceptual dedup, OCR, transcript+OCR search, the wall-clock ladder, MCP tools with embedded usage examples, six workflow prompts, and a findings contract. One uvx command instead of an afternoon of glue.

Is it really local? What leaves my machine? Nothing goes up, ever. For local files the network is used only for one-time downloads (ffmpeg build, whisper/OCR/diarization models). process_url is the single tool that talks to the network at runtime, and only down: it fetches the public source you named, once. No telemetry. See Privacy — and SECURITY.md treats a violation of this promise as a vulnerability.

For agents & tooling

Machine-readable entry points, so AI agents can install and use this server without a human reading docs:

• llms-install.md — step-by-step install instructions for agents
• llms.txt — index of the documentation
• .agents/skills/talkthrough/SKILL.md — an Agent Skill teaching the tool workflow; discovered automatically inside a checkout by Codex CLI ($talkthrough) and readable by Claude Code, Cursor, Copilot, Gemini CLI and other SKILL.md-compatible tools
• AGENTS.md — instructions for coding agents contributing to this repo
• server.json — MCP registry manifest
• integrations/ — per-engine adapters, all generated from one source of truth and drift-tested (incl. the Claude Code plugin under integrations/claude-code/)
• docs/URL_ACCEPTANCE_CORPUS.md — the live URL corpus behind the release QA of process_url (manual, needs the network; CI stays offline)

Roadmap

cloud STT · embeddings/semantic search · hosted/remote mode · .mcpb bundle · whisper.cpp backend

License

MIT