Espresso Horoscope MCP
Local-first MCP project that turns espresso shot metrics into a personalized cosmic reading, generated offline with GPT-OSS through LM Studio.
Portfolio snapshot
What it demonstrates: local model inference, MCP-style tool boundaries, structured sensor data prompting, deterministic generation, and a lightweight longitudinal memory pattern.
User experience: enter a birth date, use simulated or recorded espresso shot data, and receive a tarot-style horoscope card that combines brewing diagnostics with GPT-OSS-generated prose.
Why it exists: I wanted to learn how to use local open-weight models under real pressure. The best way for me to do that was to enter a time-boxed hackathon, keep the scope realistic, and build something weird enough to stay interesting. Plus, I want to turn technical data like coffee brewing data (profile, pressure, flow rate, etc.) into something fun that everyday coffee drinkers can relate to and learn from.
Why this exists
Built for the OpenAI Open Model Hackathon, Best Local Agent category. The category asked for useful agentic applications of GPT-OSS that could run without internet access.
The honest learning goal was simple: get my hands dirty with local model inference. A weekend hackathon with a real submission deadline gave me the right kind of constraint. It forced pragmatic choices, scope cuts, and a working demo instead of endless tutorials.
The horoscope framing is deliberately playful. The architecture underneath it is the real point: a reproducible, offline-first local agent workflow that combines structured domain data with local LLM inference through an MCP-shaped interface.
This is not fortune-telling. It’s a compact demo of how a local agent can take real-world machine data, preserve deterministic behavior where needed, and still produce an expressive user-facing result.
What it does
Connects espresso shot data to a local LLM and generates a short personalized reading from:
- Brew ratio, pressure, temperature, flow rate, extraction time, channeling signals
- Birth date for the cosmic framing layer
The architecture enforces a strict tool/prompt boundary between deterministic sensor parsing and the local generative model.