The Model Context Protocol (MCP) lets Claude, Cursor, and other AI tools query external data sources directly. Most people use it for databases or code repos. Burn 451 uses it for your saved articles — so you can ask Claude "what did I save about protein folding last month" and get answers from your own reading history.

How Burn 451's MCP server works

When you connect to Burn 451's MCP endpoint, Claude gets three tools: search_articles (semantic search over your saves), get_article (fetch full content by ID), and list_recent (chronological feed with metadata). The server runs server-sent events over HTTP — no local daemon needed.

Setup in Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "burn451": {
      "command": "npx",
      "args": ["-y", "@burn451/mcp-server"],
      "env": {
        "BURN451_API_KEY": "your-api-key-here"
      }
    }
  }
}

Build your own MCP read-later server

If you use a different read-later tool or want full control, here's the pattern:

Choose your storage — SQLite for local, Supabase/PG for cloud. You need title, URL, content text, saved_at, and an embedding column for semantic search.
Generate embeddings on save — OpenAI text-embedding-3-small or a local model. Store alongside the article.
Expose MCP tools — use the @modelcontextprotocol/sdk TypeScript package. Define tools with Zod schemas. Return JSON.
Add semantic search — on search_articles call, embed the query, cosine-similarity against stored embeddings, return top-k.

The full pattern is covered in 3 patterns from building Burn 451's MCP server. The short version: the hard part isn't the MCP protocol, it's the content extraction and embedding pipeline.

See also: Best Chrome Bookmark Extension 2026 — the browser-side complement to an MCP server.