Kay Chen  /  GitHub repo

A personal AI system for training, cooking, and eating out

I kept running into the same problems:

  • Food going bad before I used it, while I bought more for the next recipe.
  • Recipes scattered across RedNote, NYT Cooking, blogs, and screenshots.
  • Workout history that wasn't there when I needed it at the gym.
  • Over a thousand saved restaurants, and still no quick answer to "where should we eat?"

So I built an AI agent around my actual routines. I use it every day to keep my kitchen inventory current, choose what to cook, save recipes, run my workouts, and decide where to eat.

9
workouts
147
sets logged
14
meals logged
95
items in my kitchen
228
saved recipes
1,492
saved restaurants

From my live databases, as of Oct 5, 2026

Design rule If two runs of the model could disagree on something, it belongs in code. The model makes judgment calls. Facts live in databases, deterministic logic lives in the servers, procedures live in skills, and rules live in one versioned file.

See it work

Four things I actually say to it. Step through one to see which skill and which database handle each part.

    Interfaces

    Project: rules and routing

    Skills: procedures

    MCP servers: facts and logic

    Outside services

    How it evolved

    I only build when something breaks or becomes a weekly pain. Each change starts from something I noticed in daily use, and each one has a list of what I deliberately didn't build.

      How it's built

      Built on Claude with Model Context Protocol (MCP) servers I wrote and deployed on Render. Everything has exactly one home.

      MCP servers

      Facts, schema, and deterministic logic: dates, sums, diffs, queries.

      Would two runs of the model disagree? Put it in code.

      Skills

      Reusable procedures and general knowledge, like how to pick a recipe or log a meal.

      Useful in any project, or shareable with someone else.

      Project instructions

      Role, coaching rules, and which skill handles what.

      Specific to how I'm coached.

      Chat memory

      Almost nothing.

      A fact goes in a database; a rule goes in a file.

      The three servers

      kitchen-inventory

      8 tools

      Items by storage location with quantity, expiry, and low-stock thresholds. Computes what's expiring and what's running low.

      workout-log

      19 tools

      Program, sessions and sets, exercise history and form notes, cardio and race plans. Computes today's training status and weekly totals.

      food-journal

      26 tools

      Home cooking, restaurant visits and dishes, saved places and recipes, the "this week" list, and a shared dish catalog that joins cooking and eating out.

      Design decisions I'd defend