Mcp

A month ago I wrote about building a self-hosted memory layer for Claude Code. The system has been running continuously since then — 194 memories stored, mostly from Claude Code sessions across a dozen projects. This post covers what I learned from actual usage and the changes I made as a result. What Changed After Real Use The original system worked. Search found relevant memories, AI metadata extraction added useful structure, and the mobile capture form let me save thoughts from my phone. But a month of daily use exposed friction points that weren’t obvious during initial development. ...

I’ve been running Claude Code as my primary development tool for months now. It writes code, reads docs, manages git — all from the terminal. But there’s always been a gap: it can’t see my email, my calendar, or my contacts. If I ask “what’s on this week?” it has to guess. If I need to draft a reply to someone, I’m switching to the browser. Fastmail MCP Server fills that gap. MCP (Model Context Protocol) is a standard that lets AI tools call external services through a uniform interface — the AI discovers available tools, calls them with structured inputs, and gets structured outputs back. An MCP server is just a process that exposes those tools. This one connects to Fastmail’s JMAP API and gives any MCP client access to 38 tools across email, contacts, and calendar. ...

Most AI tools have some form of memory now — Claude Code has its CLAUDE.md files, ChatGPT remembers things between sessions, Cursor has rules files. But these memory systems are siloed to one tool, stored as flat text, and not searchable by meaning. You can’t query “what did I decide about authentication last month?” and get a useful answer. Your context doesn’t travel between tools, and there’s no structure beyond what you manually write. ...