# korovin-aa97/talkthrough-mcp [Health: Active]

**Category:** 🎥 Multimedia Process  
**Repository:** https://github.com/korovin-aa97/talkthrough-mcp  
**GitHub Stars:** 26  
**npm Downloads (last month):** 7778358  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/korovin-aa97-talkthrough-mcp

## Description
Turn narrated screen recordings (or any local video/audio) into agent-ready data: local Whisper transcript, scene keyframes, OCR, full-text search, and wall-clock anchoring that maps spoken remarks to real timestamps for log correlation. No cloud, no LLM inside. uvx talkthrough-mcp.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "talkthrough-mcp": {
    "command": "uvx",
    "args": ["--python"]
  }
}
```

## Documentation

## What korovin-aa97/talkthrough-mcp MCP server does

The korovin-aa97/talkthrough-mcp MCP server converts video and audio into structured, agent-ready evidence. Its processing pipeline can produce a local Whisper transcript, scene keyframes, OCR text, full-text search data, and wall-clock anchoring. Wall-clock anchoring connects remarks in the recording to real timestamps, which supports comparing spoken events with logs or other time-based records.

Inputs can be local `.mov` and `.mp4` files, other local video or audio formats, and public video links when the URL extra is installed. The project describes workflows for narrated bug recordings, meetings, workshops, product demonstrations, and production incidents. It does not post issues or send processed media to a tracker; tracker actions remain the responsibility of the surrounding agent workflow and its own tools.

## How it works

The korovin-aa97/talkthrough-mcp MCP server runs as a local stdio MCP process. An agent sends media-processing and evidence-search requests through MCP, while the server performs transcription, frame extraction, OCR, and indexing on the local machine. The resulting evidence can then be searched or used to assemble material such as an issue draft.

Whisper models download on first use, and media or model assets may require network access during initial setup. Subsequent jobs can reuse shared Whisper, OCR, and diarization caches. The README also notes that a bundled FFmpeg build can be fetched when a system FFmpeg installation is unavailable.

## Setup and configuration

The documented runner is `uvx`, with Python constrained to versions 3.11 through 3.13. A direct setup command is:

```bash
uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]"
```

The `diarization` extra enables the installation variant shown in the README, while `url` supports public video-link workflows. The only stated prerequisite for the quickstart is `uv`; the project says FFmpeg falls back to a bundled build and that OCR is installed through Python packages.

The server can be registered in Claude Desktop, Cursor, Cline, Windsurf, and other MCP clients using a stdio configuration. Claude Code also supports adding the server with `claude mcp add`, or installing the repository’s full plugin for its additional slash commands, triage subagent, and skill. The plugin and the standalone server should not both be registered in the same Claude Code setup.

## Tools and capabilities

The README describes nine MCP tools and six prompts, although the provided material does not enumerate their individual names. The supported capability areas are:

- Process local video and audio recordings.
- Accept public video links with the URL-enabled installation.
- Generate local Whisper transcript segments.
- Extract scene keyframes and exact frames for visual evidence.
- Run OCR over recorded screen content.
- Search indexed evidence using full text.
- Map spoken remarks to wall-clock timestamps.
- Support diarization through the documented installation extra.

These outputs are intended for agent investigations and evidence-backed drafts rather than for automatically filing issues. The example workflow uses the resulting evidence with separate tracker tooling.

## Limitations and notes

The korovin-aa97/talkthrough-mcp MCP server is local-first: the README states that it does not use a cloud service or an LLM inside the server. It also does not leave the machine to post issues. Agents still need separate tools if they must create tickets or interact with a source-control or issue-tracking system.

First-run processing can be slow because `uvx` resolves the environment and `process_media` may download media, Whisper, OCR, or diarization assets. Updating the plugin can create a new environment and may cause the bundled FFmpeg download to happen again. Silent recordings can produce keyframes and OCR while yielding no transcript segments because there is no audio track.

The project targets Python 3.11–3.13 and is released under the MIT license. No API key or cloud credential is described in the supplied material.

_Full upstream README: https://allmcps.com/mcp/korovin-aa97-talkthrough-mcp/readme_