video-research-mcp

A research MCP server and workflow bundle for video understanding, document analysis, cited web research, academic discovery, and optional knowledge storage. The runtime defaults to Gemini 3.8 Flash; Claude Code workflows orchestrate tools using the active session model.

Watch the full demo on YouTube

What's in the box

A Claude Code workflow bundle with 51 tools across the monorepo, 17 slash commands, 13 skills, and 7 sub-agents. The MCP servers provide the tools. The commands give you quick workflows (/gr:video, /gr:research). The skills teach Claude how to use everything -- from research to cinematic video production. The agents handle bounded independent tasks such as research and visualization. The default installer connects the published research runtime, Playwright, and MLflow; the two video companion servers require local source setup.

Install

npx video-research-mcp@latest
export GEMINI_API_KEY="your-key-here"

One command. One API key. The installer copies commands, skills, and agents to ~/.claude/ and registers the published research runtime in ~/.claude.json. Set the API key in ~/.config/video-research-mcp/.env so GUI clients can load it, restart the client, then run /gr:doctor quick to inspect the active setup. Shell exports apply only to clients launched from that shell.

npx video-research-mcp@latest --check     # show install status
npx video-research-mcp@latest --uninstall  # clean removal
npx video-research-mcp@latest --local      # install for this project only

Requires Python >= 3.11, uv, Node.js >= 22, and a Google AI API key.

Source releases and registry packages are separate: npx ...@latest and uvx resolve published registry artifacts. See source onboarding to run a specific checkout, including the optional companion servers. Installation and local tests do not establish a live provider smoke.

What it does

Watch a meeting recording

/gr:video-chat ~/recordings/project-kickoff.mp4
> "Create meeting minutes in Dutch. Screenshot every shared screen."

Gemini analyzes the supplied video and returns timestamps, decisions, and action items; verify requested coverage and transcript fidelity against the original when accuracy matters. For local files, ffmpeg extracts frames at key visual moments. Files over 20MB are uploaded to Gemini's File API and context-cached -- follow-up questions reuse the cache instead of re-uploading.

Analyze a YouTube tutorial

/gr:video https://youtube.com/watch?v=...

Same capabilities, applied to YouTube. Precise timestamps, a concept map, and comment sentiment analyzed in the background.

Research a topic with evidence grading

/gr:research "HNSW index parameters for high-dimensional embeddings"

Runs web search and Gemini analysis in parallel. Every finding gets an evidence tier -- Confirmed, Strong Indicator, Inference, or Speculation -- so you know how much weight to give each claim. Results are visualized as an interactive evidence network.

Analyze papers, URLs, or directories

/gr:analyze https://arxiv.org/abs/2401.12345
/gr:analyze ~/papers/attention-is-all-you-need.pdf
/gr:analyze ~/papers/                              # cross-document comparison

Works with PDFs, URLs, and raw text. Extracts entities, relationships, and key arguments. Point it at a directory and it compares all documents in a single pass. Supports PDF, TXT, MD, HTML, XML, JSON, CSV.

Research grounded in source documents

/gr:research-doc ~/papers/
/gr:research-doc paper1.pdf paper2.pdf "Compare methodologies and find contradictions"

Four-phase pipeline: Document Mapping, Evidence Extraction, Cross-Reference, Synthesis. Every claim is cited back to document and page number.

Search the web

/gr:search "latest developments in MCP protocol"

Google Search via Gemini grounding with source citations.

Recall what you've learned

/gr:recall                                # overview: stats + saved analyses
/gr:recall "kubernetes"                   # semantic search + filesystem grep
/gr:recall ask "what do I know about X?"  # AI-powered Q&A with source citations

Analysis tools attempt write-through storage when Weaviate is configured; storage failures are non-fatal and must be checked separately. Commands also save local notes when their workflow runs. Weeks later, in a different project, you just ask. When Weaviate is configured, searches use semantic matching. Without it, recall falls back to exact keyword grep over saved files.

Use it as a standalone MCP server

The tools are standard MCP. Any MCP client can call them -- no Claude Code required.

{
  "mcpServers": {
    "video-research": {
      "command": "uvx",
      "args": ["video-research-mcp"],
      "env": { "GEMINI_API_KEY": "${GEMINI_API_KEY}" }
    }
  }
}

Commands

How a command runs

/gr:video-chat ~/recordings/call.mp4
> "Summarize this meeting, extract action items"

 Phase 1   Gemini analyzes the video
 Phase 2   Results saved to memory
 Phase 2.5 ffmpeg extracts frames (local files only)
 Phase 3   Concepts and relationships enriched
 Phase 4   Interactive visualization generated (opt-in)
 Phase 5   Playwright screenshots it
 Phase 6   Everything copied to output/<slug>/

Visualization and comment analysis run as background agents, so they don't block the conversation.

Output structure

output/project-kickoff-2026-02-28/
├── analysis.md          # timestamped analysis with YAML frontmatter
├── frames/              # extracted video frames (local files)
├── concept-map.html     # interactive visualization
└── screenshot.png       # static capture

Tools

Skills

Skills provide repeatable workflows and load when relevant or explicitly invoked. Provider guidance does not install external image, video, or TTS connectors.

Research & analysis

Media production (new in v0.6.0)

The production skills use progressive disclosure: core patterns load when triggered (~1,000 words each), detailed recipes and reference tables live in references/ and load on demand.

Knowledge store

Connect Weaviate, and everything you learn gets stored -- searchable across projects, across sessions. Without it, the plugin works the same; you just don't get persistent semantic search.

Thirteen collections are created on first connection:

Knowledge graph

Analysis tools (content_analyze, video_analyze, research_deep, research_web, research_document, content_batch_analyze) automatically extract concepts and relationships. Over time this builds a queryable knowledge graph across all your research -- no manual step required.

Eight knowledge tools query this data: hybrid search with optional Cohere reranking, semantic similarity, fetch by UUID, manual ingest, schema introspection, and collection stats. knowledge_ask uses Weaviate's QueryAgent for AI-generated answers with source citations.

# install QueryAgent support
uv pip install 'video-research-mcp[agents]'

To set up Weaviate:

export WEAVIATE_URL="https://your-cluster.weaviate.network"
export WEAVIATE_API_KEY="your-key"

Configuration

Other install methods

Standalone MCP server (no plugin assets)

{
  "mcpServers": {
    "video-research": {
      "command": "uvx",
      "args": ["video-research-mcp"],
      "env": { "GEMINI_API_KEY": "${GEMINI_API_KEY}" }
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "video-research": {
      "command": "uvx",
      "args": ["video-research-mcp"],
      "env": { "GEMINI_API_KEY": "your-key-here" }
    }
  }
}

From source

git clone https://github.com/Galbaz1/video-research-mcp
cd video-research-mcp
uv venv && source .venv/bin/activate && uv pip install -e ".[dev]"
node bin/install.js --global

Development

uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
uv run pytest tests/ -v        # 781 tests, all mocked
uv run ruff check src/ tests/  # lint

Troubleshooting

Contributing

See CONTRIBUTING.md for development setup and PR guidelines. See ROADMAP.md for planned work. Report security issues via SECURITY.md.

Author

Fausto Albers -- Lead Gen AI Research & Development at the Industrial Digital Twins Lab, Amsterdam University of Applied Sciences (HvA), in the research group of Jurjen Helmus. Founder of Wonder Why.

Credits

• Google Gemini (google-genai SDK) -- The configured Gemini models provide video understanding, structured outputs, search grounding, and context caching; feature support is checked per model.
• FastMCP -- MCP server framework. The composable sub-server pattern (app.mount()) keeps 51 tools organized across 3 servers.
• Weaviate -- vector database powering the knowledge store. Thirteen collections, hybrid search, and the Weaviate Claude Code skill that inspired the knowledge architecture.
• ElevenLabs -- text-to-speech with word-level timestamps for voiceover generation.
• video_explainer by prajwal-y -- the video synthesis engine behind the explainer pipeline.
• Pydantic -- schema validation for all tool I/O. Structured generation via model_json_schema().
• MLflow (mlflow-tracing) -- optional observability. Every Gemini call becomes a traceable span with token counts and latency.
• Cohere -- optional reranking in knowledge search for better result relevance.
• Remotion -- React-based video rendering for the explainer pipeline.
• Claude Agent SDK -- powers parallel scene generation in video-agent-mcp.

License

MIT