Video Reader MCP

Evidence-first video reading for AI agents — ffprobe, subtitles, scenes, transcripts, and timelines without frame-by-frame LLM vision.

Status: v0.1.0 shipped — read_video MCP tool available.

Orchestrated by smart-reader-mcp — portfolio ADR lives there, not in pdf-reader-mcp.

Read vs interpret

Read (this repo): extract facts, metadata, transcripts, regions, and timelines with provenance — no generative LLM required.

Interpret (out of scope): summarize, classify, or answer open questions — belongs in the agent or an optional remote provider adapter.

Related MCP server: vidlizer

MCP surface

Primary tool: read_video

Returns a timeline document per source:

• ffprobe format + stream metadata

• chapter markers

• embedded subtitle cues (time_ms, text, provenance)

• optional scene boundaries (ffmpeg scene filter)

• warnings (missing ffmpeg/ffprobe, VFR, missing audio, skipped ASR)

• optional local ASR hook (skipped in v0.1 unless adapter is wired)

No per-frame vision LLM calls. No cloud APIs by default.

Prerequisites

• Node.js ≥ 22.13

• ffprobe (required) and ffmpeg (recommended for subtitles + scenes) on PATH

Quick start

npx @sylphx/video-reader-mcp

From source:

bun install
bun run build
bun run start

Example read_video input

{
  "sources": [{ "path": "./sample.mp4" }],
  "include_subtitles": true,
  "include_scenes": true,
  "scene_threshold": 0.4
}

HTTP transport (optional)

MCP_TRANSPORT=http MCP_HTTP_PORT=8080 node dist/index.js

Development

bun install
bun run typecheck
bun test
bun run build

Unit tests mock parsers and do not require ffmpeg in CI. Integration with real media is optional locally.

License

MIT © SylphxAI