Local Agentic Horticulture Tracker 🌿

An autonomous, local-first AI agent designed to track indoor horticulture metrics, including cold-stratification timelines, ericaceous soil pH validation, and transplant-shock recovery.

This project demonstrates the implementation of a cyclic LangGraph reasoning engine communicating with a FastMCP tool server, powered entirely by a local offline LLM.

🧠 System Architecture & Workflow

The agent operates on a continuous reasoning loop, evaluating user input to determine if external tool execution is required before generating a final response.

Workflow Breakdown:

1. Receive User Input: The human query is passed to the LangGraph state machine.

2. LLM Reasoning (The Brain): The local model (Qwen2.5) evaluates the context against the system prompt.

3. Tool Call Execution: If the LLM determines a tool is needed (e.g., logging stratification), it pauses generation, triggers the FastMCP server, and waits for structured data.

4. Context Update: The tool's output is fed back into the reasoning loop.

5. Generate Final Response: Once all necessary data is gathered, the LLM synthesizes a final, human-readable response.

Related MCP server: Garmin MCP

🛠️ Tech Stack

• Framework: LangGraph (State Machine / Agent Loop)

• Tooling Protocol: FastMCP (Model Context Protocol)

• Local LLM: Ollama (qwen2.5)

• Language: Python 3.11+

• Validation: Pydantic (Strict schema enforcement)