# Talkthrough

> Analyze narrated screen recordings and audio files through the talkthrough MCP server — triage feedback into findings, extract specs/backlogs/action items from recordings, and correlate spoken remarks with logs via wall-clock timestamps. Use when the user mentions a screen recording, screencast, narrated video/audio file, or asks to "watch" a recording and act on it.

- **Type:** Skill
- **Install:** `agentstack add skill-korovin-aa97-talkthrough-mcp-talkthrough`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [korovin-aa97](https://agentstack.voostack.com/s/korovin-aa97)
- **Installs:** 0
- **Category:** [Content & Media](https://agentstack.voostack.com/c/content-and-media)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [korovin-aa97](https://github.com/korovin-aa97)
- **Source:** https://github.com/korovin-aa97/talkthrough-mcp/tree/main/integrations/claude-code/skills/talkthrough

## Install

```sh
agentstack add skill-korovin-aa97-talkthrough-mcp-talkthrough
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Analyzing narrated recordings with talkthrough

The talkthrough MCP server turns a local video/audio file into queryable
structured data: timestamped transcript segments, scene keyframes, OCR'd
on-screen text, and wall-clock anchoring. No LLM inside — you bring the
reasoning; it brings the evidence. Everything is lazy and token-budgeted:
never ask for more than the moment you are analyzing.

## Prerequisite

The `talkthrough` MCP server must be connected (tools like
`process_media` / `get_transcript` are visible). If not, tell the user to
install it: `claude mcp add -s user talkthrough -- uvx talkthrough-mcp`
(see the repository README for other clients).

## Core workflow

1. **Ingest once**: `process_media(path)` — idempotent by content hash;
   re-calls on the same file return instantly. Long videos take minutes and
   stream progress. The summary gives you `job_id`, counts, wall-clock, and
   a transcript preview — do NOT dump anything else eagerly.
2. **Orient**: `get_transcript(job_id)` (paginate via `next_start_ms` when
   `truncated`) or `search(job_id, "")` to jump straight
   to the relevant moments (searches speech AND on-screen OCR text).
3. **Evidence per remark**: `get_moment(job_id, t0-2000, t1+2000)` — one
   call returns the transcript slice + up to 3 unique frames + their OCR
   text + the wall-clock range. This is the workhorse; describe `observed`
   from the returned pixels, never from imagination.
4. **Precision when needed**: `get_frames(at_ms=...)` for nearby keyframes;
   `extract_frame(job_id, at_ms, crop={x,y,w,h})` for an exact instant at
   native resolution (keyframes capture scene changes + a 1 fps floor, so
   sub-second moments can fall between them).
5. **Recall across sessions**: `list_jobs()` — the store persists; a file
   processed yesterday (even via CLI) is queryable by `job_id` today.

## Timestamps

Every timestamped result carries `t_ms` (video-relative) and, when the
recording start is known, `t_wall` (ISO 8601 real time). Use `t_wall` to
correlate remarks with server/app logs (±30 s grep window). If
`wall_clock` is null or low-confidence, ask the user when the recording
started and re-anchor: `process_media(path, recorded_at="",
force=true)`.

## Packaged workflows (server prompts)

Prefer the server prompts when the task matches — they encode the full
method: `triage-recording` (screencast → findings JSON per the contract in
`examples/output-contract.schema.json`), `spec-from-workshop`,
`backlog-from-demo`, `meeting-actions` (audio-only friendly),
`correlate-with-logs`.

## Rules of thumb

- Audio-only jobs (.m4a/.mp3/…): transcript tools work; frame tools error
  by design — that error is expected, not a failure.
- Findings/quotes must cite the narrator's exact words + `t_ms` (+ `t_wall`
  when known) + the frame files you actually inspected.
- Low STT/vision confidence → surface a question; never silently guess.
- Any narration language works (Whisper auto-detects; the summary reports
  `language` + `language_probability`). Garbled transcript or low/wrong
  detection → re-call `process_media(path, model="large-v3-turbo",
  force=true)` (best multilingual quality) or pin `language="…"`; domain
  jargon → pass `vocabulary="Term1, Term2"`.
- Write digests/summaries for the recording author in the narrator's
  language; keep quotes verbatim in the original — translate in your own
  prose only, never inside a quote.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [korovin-aa97](https://github.com/korovin-aa97)
- **Source:** [korovin-aa97/talkthrough-mcp](https://github.com/korovin-aa97/talkthrough-mcp)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-korovin-aa97-talkthrough-mcp-talkthrough
- Seller: https://agentstack.voostack.com/s/korovin-aa97
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.