For humans & agents
MCP Server
Why I built this
Every personal site is a content silo. An AI agent asked about my work has to scrape, summarise, and probably hallucinate. I didn't want that. So I gave the site a machine-readable surface — an MCP server, running on Cloudflare Workers, speaking the same protocol Claude Desktop and every agent framework now speaks natively.
Now when someone's AI assistant asks "who is Naren Katakam and what has he built?", it gets a structured answer from the source — not a degraded guess.
Tools
get_person() Profile, expertise, and primary contact channels.
e.g. "Who is Naren and what does he focus on?"
search_projects(query) Semantic search across case studies.
e.g. "Find his work on cold-start recommendation engines."
ask(question) RAG-backed Q&A over all site content + the Vedantic AI paper.
e.g. "What is the Data Trinity framework?"
traverse_graph(concept) Walk the knowledge graph from any concept.
e.g. "Show me everything connected to 'pramana'."
Connect in Claude Desktop
Add this to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) and restart Claude.
Config
{
"mcpServers": {
"narenkatakam": {
"url": "https://mcp.narenkatakam.com/mcp",
"transport": "streamable-http"
}
}
}
Architecture
- Cloudflare Workers — streamable HTTP MCP server (~50 lines).
- D1 — SQLite for facts + structured content.
- Vectorize + Workers AI — embeddings + semantic search.
- RAG fallback — FTS5 full-text when vector search returns low confidence.
- Six-pramana tagging on every answer (pratyakṣa = retrieval, śabda = source, anupalabdhi = absence check).
Steal this pattern
The whole thing is roughly: one Worker, one D1 database, one Vectorize index, one Custom Domain. Total cost: $5/month (Workers Paid plan). If you're building a personal site in the age of agents, give it an agent interface. The pattern generalises — any knowledge-heavy surface benefits.
Not open-source yet. Might be, once TrustCitation ships. If you want the skeleton, DM me.