Contendeo

Contendeo

Give your agent eyes.

Contendeo is the multimodal layer that lets your AI actually see video — not just read its transcript.

---

What it is

Contendeo is a remote MCP server that gives LLMs frame-level context from video — transcription, keyframe vision analysis, OCR, and structured output, unified into a single response your agent can reason over.

Paste a YouTube, Instagram Reels, Vimeo, Twitter/X, TikTok, or direct video URL into Claude (or any MCP client). Contendeo downloads, transcribes, extracts keyframes, runs OCR, analyzes visuals, and returns a structured document.

Live at [contendeo.app](https://contendeo.app). MCP endpoint at [contendeo.app/mcp/](https://contendeo.app/mcp/).

Why it exists

Transcripts capture what was said. They don't capture what was shown — chart values, UI states, overlays, code on screen, circled regions, dashboard numbers. For any video where the visual layer carries information (trading tutorials, product demos, technical walkthroughs, data dashboards, design reviews), transcript-only analysis misses half the signal.

Contendeo closes that gap.

See the side-by-side comparison at [contendeo.app/demo](https://contendeo.app/demo).

Tools

Four tools exposed over MCP. Credit costs charged per successful call. Cache hits are free.

Full schemas and response formats: docs/tools.md.

Installation

Contendeo is a remote MCP server — no local install, no package download. Connect via URL from your MCP client.

Claude Desktop

Add to your claude_desktop_config.json:

{

"mcpServers": { "contendeo": { "url": "https://contendeo.app/mcp/" } } }

Claude will walk you through OAuth on first use.

Claude Code

claude mcp add --transport http contendeo https://contendeo.app/mcp/

Other MCP clients (Cursor, Continue, Cline, etc.)

Any client that supports remote MCP servers over streamable HTTP. Point it at https://contendeo.app/mcp/.

Full walkthrough with screenshots: docs/installation.md.

Authentication

Contendeo uses OAuth 2.0. First use flow:

1. Your MCP client initiates a connection
2. Browser redirect to contendeo.app sign-in (Google OAuth or email)
3. Approve the MCP client
4. Bearer token issued, client reconnects, tools become available

New accounts get 10 free credits. No card required.

Details: docs/authentication.md.

Pricing

Cache hits are free. Failed analyses auto-refund.

Details: docs/pricing.md.

Stack

Under the hood:

• [yt-dlp](https://github.com/yt-dlp/yt-dlp) — video download across 1000+ sites
• [Apify](https://apify.com) — Instagram Reels fetching
• [Webshare](https://webshare.io) — residential proxy pool for YouTube anti-bot
• [ffmpeg](https://ffmpeg.org) — audio extraction + keyframe scene detection
• [Groq Whisper](https://groq.com) (large-v3) — transcription
• [Tesseract OCR](https://github.com/tesseract-ocr/tesseract) — deterministic text extraction from frames
• [Claude Vision](https://www.anthropic.com/api) (Sonnet 4.6) — visual frame analysis
• [Supabase](https://supabase.com) — auth + Postgres
• [Redis](https://redis.io) — URL-hash result cache
• FastAPI + MCP Python SDK — server, streamable HTTP transport
• Traefik — SSL termination + routing

Links

• Landing — contendeo.app
• Demo — contendeo.app/demo
• MCP endpoint — contendeo.app/mcp/
• Dashboard — contendeo.app/dashboard
• Docs — docs/
• Server card — contendeo.app/.well-known/mcp/server-card.json

Support

Questions, bug reports, feature requests: open an issue.

Commercial/partnership inquiries: @0xKaroshi on X.

License

MIT — covers this wrapper repo (documentation, examples, manifest). The production server source is not open-sourced.

See LICENSE.

---

*Contendeo — every frame, every word, every edge.*