Case study · MCP Servers
Three MCP servers, so an AI assistant can see my actual day.
MCP servers are how an AI app gets tools. I built three, in Python, for the services I actually live in: Google Workspace, TickTick and Rebrickable. They’re how my assistant reads my calendar and task list, logs a job application in a spreadsheet, and tells me to go to bed without being asked.
The problem
An AI that can only see the conversation can’t help with the part of ADHD that costs me anything: the thing I meant to do, the deadline I couldn’t feel coming, the fact that it’s one in the morning and nobody has said so. To be any use, it needed to see my real life, and there were no good tools for the services I use.
Why I built it
I wanted an assistant that notices things without being asked. Now I mention that I applied for a job and a row appears in my applications sheet, and when it’s late, it tells me to go to bed.
How I built it
- Google Connections. 107 tools across Gmail, Calendar, Drive, Docs, Sheets and Tasks, plus read-only Search Console and Google Analytics. Every tool is named for its service and marked read-only or not, so an AI picks the right one and knows which calls change anything.
- A real sign-in. The server speaks Streamable HTTP behind a full OAuth 2.1 authorization server: discovery, dynamic client registration, PKCE, rotating refresh tokens and revocation. Google is the login, and only my account gets in. Tokens are stored hashed on a persistent volume, so the connection survives a redeploy, and the old open endpoint switches itself off once the new sign-in has been used.
- TickTick, past its public API. I started from an open-source TickTick server built on the official API and extended it well past that: I worked out TickTick’s unpublished endpoints from the app’s own traffic to add task pinning, repeat-from-completion, subtasks, checklist items and activity history, none of which the documented API offers, and rebuilt its sign-in so it stops tripping TickTick’s login limits.
- Rebrickable. LEGO parts, colors and part lists, because a hobby is a real problem too.
- The same pattern in a product. Pockets by Hoopy ships its own MCP server with the same kind of OAuth 2.1 sign-in, so its users can connect Claude or ChatGPT to their budget. Read the Pockets case study.
Stack
- Python
- MCP SDK
- FastMCP
- OAuth 2.1
- Google APIs
- Railway