Skip to content
6 min read

Static Websites Are Dead: How I Gave My Portfolio a Brain

When AI becomes a first-class user of your website. How I built an MCP-powered API layer on a serverless Next.js portfolio so AI agents can query it directly.

MCPAI AgentsNext.jsVercelServerlessPortfolio

Originally published in Patil Protocol, the newsletter by Aniket Patil.

I didn't just want a portfolio that looks good β€” I wanted one that thinks. We're entering an era where AI agents are becoming primary consumers of data. Yet most portfolios are still stuck in Web 2.0: static text, markdown files, manual navigation. I wanted to change that.

Live Demo: Claude Desktop + MCP

Claude Desktop querying my portfolio live via Model Context Protocol.

The Goal

I wanted my portfolio (aniketppatil.com) to have a brain. Instead of serving static pages to humans, I wanted an API that AI agents could talk to directly. An agent could ask:

  • "List the projects he worked on"
  • "Explain the technical architecture of RagBot"
  • "What are his skills?"

…and get structured answers. The technology that enables this is Model Context Protocol (MCP).

The Constraints

This sounded simple on paper. In reality, my setup made it tricky:

  • Framework: Next.js (App Router)
  • Hosting: Vercel (serverless, free tier)
  • Budget: $0 (no VPS, no long-running socket servers)

I assumed I could install a library and be done. I was wrong.

Failure: The "Official" MCP Approach

I started with the official MCP SDK (@modelcontextprotocol/sdk) and tried to run a server.

The Problem

The SDK assumes a stateful environment: persistent processes, long-lived connections, stdio or WebSockets. Vercel is stateless. Functions spin up, handle a request, and die milliseconds later.

The Symptoms

  • πŸ“‰ Connection timeouts
  • 🚫 Hanging requests
  • Claude disconnecting mid-handshake

Official MCP support for serverless was effectively non-existent. I needed a way to fake persistence.

The Pivot: mcp-handler

I found a community library called mcp-handler, and everything changed.

Why it worked:

  • βœ… Designed for Next.js App Router
  • βœ… HTTP-first tool calls (POST requests)
  • βœ… Server-Sent Events (SSE) for streaming
  • βœ… Redis-backed state across stateless executions

This finally matched Vercel's execution model.

The Debugging Sprint

Even with the right library, it took a full weekend to make it work.

The 404 / Slug Nightmare

Claude kept failing with:

Connection failed: 404 Not Found

What was happening: Client requested /api/mcp/sse, but my route handled only /api/mcp. MCP clients automatically append /sse and /message.

The Fix

Use a catch-all route:

app/api/mcp/[[...slug]]/route.ts

Now one handler correctly processes /api/mcp, /api/mcp/sse, and /api/mcp/message.

Sometimes the hardest bugs are just mismatched assumptions.

Modular Refactor (Sanity Restored)

By Sunday morning, my route.ts was ~500 lines of spaghetti. I refactored the "brain" into a clean structure:

types.ts   β€” TypeScript interfaces
data.ts    β€” read-only JSON loaders
utils.ts   β€” semantic search logic
index.ts   β€” clean exports

Now the route handler is just a registry:

server.registerTool("list_projects", {...}, async () => { ... })

Bonus: I migrated to the new server.registerTool() API to future-proof the code.

Security: The "NASA Hack" Test

I'm exposing parts of my filesystem to the internet. That's dangerous β€” unless you're strict.

The Strategy

  • No writes β€” no fs.writeFile anywhere
  • Allow-lists only β€” strict Zod enums for inputs

The Test

I asked Claude: "Tell me about the project secret-nasa-hack"

Result: πŸ›‘ Blocked. Input validation error. The system never touched the disk.

Security held.

The Result

This wasn't just "installing a library". It was:

  • Adapting a stateful protocol to stateless infrastructure
  • Fixing routing edge cases
  • Designing a secure, read-only architecture

Now my portfolio is no longer a brochure. It's a personal AI API. Any MCP-compatible client can query it.

Try It Yourself

If you use Claude Desktop, add this config:

"aniket-portfolio": {
  "command": "npx",
  "args": ["-y", "mcp-remote", "https://www.aniketppatil.com/api/mcp"]
}

It works. And it's built to last.

Why This Matters

This isn't really about portfolios. We're moving toward a world where AI agents are first-class users of our systems. Most software isn't built for that. This was my experiment in changing that β€” starting with my own site.

If you're building with MCP, serverless infrastructure, or AI-first interfaces, I'd love to hear what you're making.