# Video Reader MCP

**Category:** 🎥 Multimedia Process  
**Repository:** https://github.com/SylphxAI/video-reader-mcp  
**GitHub Stars:** 3  
**npm Downloads (last month):** 21445  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/video-reader-mcp

## Description
Evidence-first video MCP. Timeline document with ffprobe streams and subtitle cues.

## Tools
Capabilities this server exposes over MCP:

- **inspect** — Document and media evidence when Markdown is not enough. video_timeline (bounded scene/cue timeline) and render_frame (actual decoded frames), and cite_check (quote/location support, not semantic truth), are anymd Pro; PDF operations: inspect (page facts, metadata), render_page (PNG images), extract_regions (crop bounding boxes), ocr_pages / analyze_regions (configured OCR or vision provider), structure (JSON with document map, elements, geometry; profile quality|research adds trust and accessibility reports), compare (page-level diff of sources[0] vs sources[1]).
- **outline** — Return a document heading tree with stable node ids, title paths, page/slide/chapter ranges and Markdown byte ranges. PDF bookmarks or detected headings, Word headings, PowerPoint slides, EPUB chapters and headings, HTML and Markdown. format json (default) or tree. Pass a node id to read to fetch that section.
- **read** — Read any document as clean Markdown: PDF, Word (DOCX), PowerPoint (PPTX), Excel (XLSX/XLS/ODS), CSV, EPUB, HTML or a web URL, Markdown/text, images (metadata + OCR), audio/video (metadata, chapters, subtitles). Pages/slides/sheets carry <!-- page N --> style markers for citation. Long documents stop at max_tokens (default 20000) and end with a cursor to continue; choose pages with pages: "1-5,8". A directory returns its readable files.
- **search** — Search documents for text: one file, many files, whole directories (recursive, .gitignore aware), or URLs, across every format read supports. Returns each hit as file + page/slide/sheet + a snippet with the match in bold. mode auto (default) finds the exact phrase and falls back to BM25-ranked passages when there is none; literal or ranked force one. Narrow directories with glob, e.g. "*.pdf".

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

```json
"mcpServers": {
  "video-reader-mcp": {
    "command": "npx",
    "args": ["-y","@sylphx/anymd"]
  }
}
```

## Documentation

## What Video Reader MCP MCP server does

Video Reader MCP works with local video files and returns structured information that an agent can use to locate and cite material. A default read includes container metadata, stream information, chapters, and embedded subtitles. Results can include timestamp values, stream indexes, subtitle ranges, and a source hash so an answer can point back to the input evidence.

The server is designed for questions such as finding a quoted line, identifying the chapter where a change occurs, or locating a useful moment in a recording. It does not create a transcript from audio. Subtitle-based answers depend on subtitles being embedded in the video; missing subtitles are reported as a gap rather than inferred.

## How it works

Video Reader MCP uses a fast default profile for initial inspection. This profile does not detect scenes, extract frames, or run speech recognition. The quality profile adds ffmpeg scene detection, but it still does not extract frames or transcribe speech.

A typical workflow starts with `read_video`, providing an absolute local path in the `sources` input. An agent can then call `search_video` to find matching embedded subtitle cues. Once a relevant timestamp is available, `video_evidence` can render or crop a frame, or apply OCR to that frame. Frame indexes are only available after this evidence step.

The server keeps scene detection, frame work, and OCR separate from the initial read. This produces a timeline first and lets an agent request visual evidence only when it has a useful timestamp. Video Reader MCP does not require a cloud video API or a frame-by-frame vision model.

## Setup and configuration

Run the released command with:

```bash
npx -y @sylphx/cue
```

Claude Code can register it as a stdio MCP server with:

```bash
claude mcp add cue -- npx -y @sylphx/cue
```

The repository README states that this project was merged into `anymd` on September 25, 2026 and is archived. The documented command above identifies the released Cue package associated with this repository.

The input uses a `sources` array containing local absolute paths. Omitting `profile` selects `fast`; setting `profile` to `quality` enables scene detection. `include_keyframes` and `include_transcript` remain unsupported in the shipped server and return a warning with an empty array when enabled.

## Tools and capabilities

- `read_video`: Reads metadata, streams, chapters, and embedded subtitles; scene detection is optional.
- `search_video`: Searches embedded subtitle cues only and does not perform speech recognition.
- `video_evidence`: Renders, crops, or OCRs one frame at a supplied timestamp.

Video Reader MCP is useful for meeting recordings, demos, and other local videos where an agent needs timestamp-level references. It can report warnings and gaps, including unavailable `ffprobe` data or absent embedded subtitles, instead of filling in unsupported information.

## Limitations and notes

The server does not run speech recognition, generate a transcript, or automatically extract frames during the default read. Quality mode adds scene boundaries but not visual frame extraction. Subtitle search works only when the source contains embedded subtitles. OCR is limited to the frame requested through `video_evidence`.

The project is licensed under MIT.

_Full upstream README: https://allmcps.com/mcp/video-reader-mcp/readme_