How I Built an MCP Server from Scratch

The Model Context Protocol (MCP) is quickly becoming the standard for how AI coding tools interact with external systems. When I started building ShackleAI, I needed persistent memory across coding sessions — something that didn't exist yet. So I built it from scratch. ## Why Build an MCP Memory Server? AI coding assistants like Claude Code, Cursor, and Windsurf are powerful, but they forget everything between sessions. Every new conversation starts from zero. You re-explain your architecture, your conventions, your decisions — over and over. I wanted a memory layer that: - Persists across sessions automatically - Searches semantically, not just by keyword - Runs locally with zero cloud dependencies - Works with any MCP-compatible tool ## The Architecture The server is built on three pillars: **1. SQLite + sqlite-vec for storage** I chose SQLite over PostgreSQL for a critical reason: zero infrastructure. Developers shouldn't need to run a database server just to have memory in their coding tool. sqlite-vec provides vector search capabilities directly in SQLite. **2. Transformers.js for offline embeddings** All embeddings are generated locally using the all-MiniLM-L6-v2 model via Transformers.js. No API keys, no network calls, no cost per query. Your data never leaves your machine. **3. MCP Protocol for integration** The server exposes 11 MCP tools that AI assistants can call: memory_store, memory_search, memory_update, memory_delete, and more. Any MCP-compatible client works out of the box. ## What I Learned Building a developer tool that's used by hundreds of developers weekly taught me several things: **Ship small, iterate fast.** The first version had 3 tools. Now it has 11. Each addition came from real user feedback. **Zero-config is non-negotiable.** If your tool requires setup steps, most developers won't use it. memory-mcp works with a single npx command. **Semantic search changes everything.** Keyword search misses context. When a developer asks "how does auth work in this project?", semantic search finds the memory about JWT middleware even though the word "auth" isn't in it. ## Results - 1,440+ weekly npm downloads - Used by developers with Claude Code, Cursor, and Windsurf - Zero-config, zero-cost, fully offline - 11 MCP tools for complete memory management ## What's Next I'm exploring cross-project memory (sharing context between related repos), team memory (shared knowledge bases for engineering teams), and deeper IDE integrations. If you're building MCP servers or agent tools, I'd love to hear what you're working on.