video-research-mcp

Claude Code can't process video. Gemini 3.1 Pro can. This plugin bridges the two -- giving Claude access to Gemini's video understanding, multi-source research, and web search through MCP.

What's in the box

A Claude Code plugin -- not just MCP servers, but a full integration: 45 tools, 16 slash commands, 6 skills, and 6 sub-agents that work together out of the box. The MCP servers provide the tools, the commands give you quick workflows (/gr:video, /gr:research), the skills teach Claude how to use everything correctly, and the agents handle background tasks like parallel research and visualization.

Install

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

One install. One API key. The installer copies 16 commands, 6 skills, and 6 agents to ~/.claude/ and configures the MCP servers to run via uvx from PyPI.

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 >= 16, and a Google AI API key.

What it does

Watch a meeting recording

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

Gemini watches the full video and pulls out timestamps, decisions, and action items. 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. You get 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. Documents are uploaded once and reused across all phases.

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

Nothing gets lost. Every analysis and research finding is stored automatically. Weeks later, in a different project, you just ask. When Weaviate is configured, searches use semantic matching -- find "gradient descent tuning" even when you searched for "ML optimization". Without Weaviate, 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

Files are also saved to Claude's project memory for /gr:recall.

Tools

Video (4): video_analyze, video_create_session, video_continue_session, video_batch_analyze

YouTube (3): video_metadata, video_comments, video_playlist

Research (8): research_deep, research_plan, research_assess_evidence, research_document, research_web, research_web_status, research_web_followup, research_web_cancel

Content (3): content_analyze, content_batch_analyze, content_extract

Search (1): web_search

Infrastructure (2): infra_cache, infra_configure

Knowledge (7): knowledge_search, knowledge_related, knowledge_stats, knowledge_fetch, knowledge_ingest, knowledge_ask, knowledge_query (deprecated)

Project (4): explainer_create, explainer_inject, explainer_status, explainer_list

Pipeline (6): explainer_generate, explainer_step, explainer_render, explainer_render_start, explainer_render_poll, explainer_short

Quality (3): explainer_refine, explainer_feedback, explainer_factcheck

Audio (2): explainer_sound, explainer_music

Scene Generation (2): agent_generate_scenes, agent_generate_single_scene

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.

Twelve collections are created on first connection:

Seven knowledge tools let you query this data: hybrid search with optional Cohere reranking, semantic similarity, fetch by UUID, manual ingest, and collection stats. knowledge_ask uses Weaviate's QueryAgent for AI-generated answers with source citations (requires the weaviate-agents package).

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

To set up Weaviate, run the interactive onboarding or set the vars directly:

/skill weaviate-setup

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        # 540 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

• video_explainer by prajwal-y -- the video synthesis engine behind the explainer pipeline. We extended it with configurable ElevenLabs voice settings, env-based configuration, and MCP tool integration. The original repo is included as a git submodule at packages/video-explainer/.

• Weaviate -- vector database powering the knowledge store. Twelve collections, hybrid search, and the Weaviate Claude Code skill that inspired the knowledge architecture.

• Google Gemini (google-genai SDK) -- Gemini 3.1 Pro provides native video understanding, thinking mode, context caching, and the 1M token window that makes all of this work.

• FastMCP -- MCP server framework. The composable sub-server pattern (app.mount()) keeps 45 tools organized across 3 servers.

• MLflow (mlflow-tracing) -- optional observability. Every Gemini call becomes a traceable span with token counts and latency.

• Pydantic -- schema validation for all tool I/O. Structured generation via model_json_schema().

• Remotion -- React-based video rendering for the explainer pipeline.

• ElevenLabs -- text-to-speech with word-level timestamps for voiceover generation.

• Cohere -- optional reranking in knowledge search for better result relevance.

• Claude Agent SDK -- powers parallel scene generation in video-agent-mcp.

License

MIT