A meal planner with PydanticAI and Ollama

June 14, 2026

A CLI agent that turns meal descriptions into a categorized shopping list - running locally with Ollama.

How it works

You describe what you want to cook in plain English. The agent checks what you already have at home and returns a structured, categorized shopping list with only the things you actually need to buy. It runs on your machine.

graph LR A[meal description] --> B[Pydantic AI Agent] B --> C[get_inventory] C --> D[inventory.json] D --> C C --> B B --> E[ShoppingList] E --> F[CLI output]

The components

Pydantic AI is the agent framework. It handles the tool-calling loop, output validation, and retries. You define tools as plain Python functions decorated with @agent.tool - the docstring becomes the tool description the LLM sees.

Ollama runs llama3.2:3b locally via a Docker container. Pydantic AI talks to it through the OpenAI-compatible /v1 endpoint.

inventory.json is the single source of truth for what's at home. The agent gets it injected as deps - Pydantic AI's way of passing runtime data into tools without globals. You declare the type at agent construction time (deps_type=dict) and access it inside any tool via ctx.deps.

A Pydantic output model (ShoppingList) enforces structure on the LLM's response. If the model returns malformed JSON, Pydantic AI retries automatically up to 3 times before giving up.

If you want to run it yourself, the setup is in the README.