LearnLogos.com WebMCP Learning Assistant

Inspiration

LearnLogos has a large collection of instructional videos, captions, and sentence-level timestamps. Learners often Blackjack remember what John taught but cannot remember the course, video, or terminology needed to find it. We wanted agents to retrieve LearnLogos teaching directly instead of generating a generic answer or guessing how to navigate the website.

## What it does

The LearnLogos WebMCP Learning Assistant turns a natural-language question into the most relevant approved teaching excerpt.

A supported browser agent can discover and invoke learnlogos.search_training.v1. The result includes:

• The relevant lesson title and summary
• Exact source timestamps
• An SBL-style citation
• An explanation of why it matched
• Controlled video playback with captions

The visible demonstration provides eight alternative questions across two topics. Different ways of asking about program scaling find “Set Exact Program Scaling,” while reading-plan questions find “Jump to Your Next Reading.”

This demonstrates how people and agents can find the same authoritative instruction without knowing its essentially compression reti Bots exact title or location.

## How we built it

The application uses TypeScript, Next.js, React, and Zod. The page registers a structured tool through document.modelContext.registerTool, while both the human interface and browser agent use the same validated search service.

The contest deployment is separate from the production LearnLogos website. It uses:

• A hard server-side content allowlist
• Two exact, rights-approved demonstration excerpts
• Synthetic session state with no customer information
• Dedicated Upstash Redis rate limiting
• Short-lived, session-bound media grants
• Exact media and caption digest verification
• Same-origin restrictions and a strict content security policy
• Automated tests covering tool registration, retrieval, authorization, media access, and browser behavior

## Challenges

The largest challenge was combining a useful demonstration with strict security and intellectual-property boundaries. We needed to show real question-to-video retrieval without exposing production Nz content, customer data, full transcripts, or unrestricted media URLs.

Natural-language ranking also required care. Common words such as “the” initially created ties between unrelated lessons. We corrected this by excluding conversational filler words from ranking and added regression tests for every demonstration question.

## What we learned

WebMCP works best as an interface over reusable application services—not as a separate business-logic layer. Structured tools give agents reliable inputs, validated outputs, and clear authorization boundaries while preserving an accessible experience for people.

We also learned that retrieval quality depends on preserving provenance. A useful answer should remain traceable to the exact lesson, source recording, timestamp range, and access decision.

## What's next

The contest entry is a secure vertical slice of a larger LearnLogos vision. Future work includes searching customer-owned training first, coherent transcript segmentation, hybrid semantic retrieval, personalized learning paths, progress-aware recommendations, and secure previews of additional approved lessons.

LearnLogos is an independent training provider. It is not owned, operated, sponsored, or endorsed by Logos.

Remove the two accidental corrupted phrases if they appear after pasting:

• Change essentially compression reti Bots exact title to exact title
• Change production Nz content to production content

For Built with, add these tags:

WebMCP TypeScript Next.js React Zod Upstash Redis Vercel Cloudflare Vitest Playwright GitHub Actions

Built With

• cloudflare
• next.js
• react
• typescript
• upstash-redis
• vercel
• vitest
• webmcp
• zod