A retrieval engine,
spoken in protocol.

Conduit puts a real pgvector hybrid-search engine behind the Model Context Protocol — exposing Sift's 84-document knowledge base as MCP tools, resources and prompts, so any client — Claude included — can search it.

Endpoint · Streamable HTTP

https://conduit-orpin.vercel.app/mcp

Connect from Claude Desktop

Add this to claude_desktop_config.json, or open the endpoint in any MCP client — try the companion Probe inspector.

claude_desktop_config.json

{
  "mcpServers": {
    "conduit": {
      "url": "https://conduit-orpin.vercel.app/mcp"
    }
  }
}

Tools

search_knowledge_basequery · mode · category?

Hybrid, vector, or keyword search over the corpus, returning ranked documents.

get_documentid

Fetch the full text of one document by its identifier.

list_categories—

List every category in the knowledge base with its document count.

Resources

Every document is a browsable MCP resource at sift://doc/{id} — a templated resource with a live listing of all 84 entries.

Prompts

The research prompt drives a grounded, multi-query investigation of any topic against the corpus.

How it works

  MCP client ──Streamable HTTP──▶  /mcp
                                    │
                                    ├─ tools · resources · prompts
                                    └─ lib/search.ts
                                        │
                                        ▼
                            Postgres + pgvector (Neon)
                            HNSW · full-text · RRF fusion