DEV Community

BuyWhere
BuyWhere

Posted on

How We Got Listed on the Official MCP Registry — A Server Builder Checklist

#mcp #ai #webdev #tutorial

After 2 months of building our MCP server and navigating the ecosystem, BuyWhere is now officially listed on the MCP Registry. Here is exactly what worked, what did not, and a checklist for other MCP server builders.

Why the Registry Matters

The MCP Registry is the canonical discovery layer for AI agents. When Claude, Cursor, or Cline searches for servers, the registry is the first place they check — not npm, not GitHub, not Google.

Getting listed turns your MCP server from "something you can npm install" into something AI agents can discover natively.

The Checklist

1. Your mcp.json Manifest (Critical) 

{
  "mcpServers": {
    "your-server-name": {
      "command": "npx",
      "args": ["-y", "@your-org/your-package"],
      "env": {
        "YOUR_API_KEY": "$YOUR_API_KEY"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode
  • Must be at the package root
  • command and args must produce a working MCP server with zero config beyond the env var
  • Test with npx -y @your-org/your-package — if it does not work locally, registry clients will fail

2. npm Package Readiness

  • package.json keywords: The npm registry uses these for search. Include mcp, mcp-server, mcp-tools, and your domain terms. We went from 24 to 33 keywords and saw a measurable lift in discoverability.
  • README.md: Must be included in your npm tarball (check with npm pack --dry-run). An empty README = zero text for npm search indexing.
  • Version stability: The registry pulls from your latest npm version. Make sure it works.

3. GitHub Repository (Minimum)

  • Public repo with clear README
  • Link to your MCP server in the description
  • Tag it with mcp-server topic
  • Working CI badge helps (shows maintenance)

4. AI Discoverability (the hidden piece)

AI models do not browse npm. They consume structured documentation:

  • llms.txt: This is the AI-native equivalent of robots.txt. Create one at your domain root with links to your docs, tutorials, and registry listing.
  • Structured tool descriptions: Write MCP tool descriptions that include example queries — this dramatically improves how Claude/Cursor selects your tools vs. others.
  • agent.json: If you support A2A (Agent-to-Agent), host your agent card at /.well-known/agent.json.

What Worked

  1. npm SEO: Fixing our keywords field (it was typod as keywordt) and adding domain-specific terms produced a 5x download spike within 48 hours.
  2. Registry listing: Once listed, AI agents began discovering BuyWhere without manual installs.
  3. Tutorial content: Our "Build a Shopping Agent in 15 Minutes" tutorial drove 770 npm downloads in a single day.

What Did Not Work

  1. dev.to alone: We published 30+ articles. Total engagement: zero. Without community interaction (comments, reactions, follows), the dev.to algorithm does not surface content.
  2. Cold outreach: Sent 40+ personalized emails to MCP developers. Reply rate: effectively zero.
  3. Relying on GitHub stars: Stars do not equal discoverability in the agent ecosystem.

The Hard Truth

The MCP ecosystem has a discoverability problem. There are 2,600+ MCP servers on GitHub but no canonical way to browse them all. The official Registry is step one, but discoverability requires:

  • npm keywords for human developers
  • Registry listing for AI agents
  • Tutorials and docs for AI model training data
  • Community presence for social proof

Your Turn

Are you building an MCP server? What is your discoverability strategy? Drop a comment — I am genuinely curious what is working (or not working) for other builders.

BuyWhere MCP — 50M+ products, 6 markets, free API key: buywhere.ai/api-keys

Top comments (0)