MCPxel

The curated navigation & rating station for Agent Skills (MCP).

Comment

Inspiration

I hated searching for Claude Skills. As a developer, every time I needed a tool (like "database manager"), I had to search GitHub, wade through 20+ repos, and spend 30 minutes installing one only to find it was broken or just a simple wrapper. Stars ≠ Reliability. I realized the MCP ecosystem needed its own "App Store" — a place where quality is verified, not just guessed.

What it does

MCPxel is a curated Navigation & Rating Station for Agent Skills.

  1. LLM-as-Judge: We use an automated pipeline to evaluate skills on 5 dimensions (Clarity, Utility, Quality, Maintainability, Novelty).
  2. S-Tier Filtering: We assign S/A/B/C grades. S-Tier (>9.0) means "install without regrets."
  3. Role-Based Search: Instead of guessing keywords, you can find tools tailored for specific roles (e.g., Product Manager, Developer).
  4. Streamlined Access: Instant preview of skill.md and usage guides, with one-click config copying. No more jumping back and forth to GitHub. ## How I built it
  5. Core: Built with Next.js for a fast, responsive frontend.
  6. AI Power: Used DeepSeek-V3 as the reasoning engine for our "LLM-as-Judge" system to evaluate skill quality.
  7. Development: Leveraged TRAE (AI IDE) to accelerate development by 5x using its agentic workflow and context awareness.
  8. Deployment: Hosted on Vercel for zero-config CI/CD and protected by Cloudflare. ## Challenges I ran into
  9. Defining "Quality": Teaching an LLM to judge code quality objectively was hard. We had to iterate on the rubric multiple times to stop it from hallucinating high scores for empty repos.
  10. Data Consistency: GitHub READMEs are a mess. Standardizing the metadata extraction for hundreds of different repos required complex parsing logic. ## Accomplishments that I'm proud of
  11. Successfully building a pipeline that can autonomously grade skills with surprising accuracy.
  12. Creating a "Role-Based Search" that actually understands user intent better than keyword matching.
  13. Shipping a polished product in record time thanks to the TRAE + DeepSeek-V3 stack. ## What I learned
  14. AI-Native Development: Coding with an Agentic IDE (TRAE) isn't just faster; it changes how you think about problems. You become an architect, not just a typist.
  15. Curation Value: In the age of AI generation, human (or AI-assisted) curation is more valuable than creation. There's too much noise; the value is in the filter. ## What's next for MCPxel
  16. Community Voting: Allowing users to upvote/downvote to complement the LLM ratings.
  17. Direct Integration: One-click install directly into Claude Desktop (once APIs allow).
  18. More Roles: Expanding the role-based search to non-technical fields like Writing and Marketing.

Built With

Share this project:

Updates