The problem. Deciding what to eat and actually getting it made is harder than it should be. Recipe search ignores what's already in your kitchen, ignores what you can't or won't eat, and leaves you alone once you start cooking.
What TALC does. TALC is a cooking and meal-planning agent for home cooks. You talk to it, and it works on real tables: your pantry, a meal plan, a shopping list and a step-by-step cooking view. It finds recipes in a collection of ~20,600 real recipes and never invents one; "not found" beats a recipe that only looks real. It plans meals around what you have, builds the shopping list from what's missing, puts what you bought into your pantry, and answers questions while you cook, saving each answer as a tip on that recipe step.
How it works. Every table can be worked by clicking or by writing: "add 3 tomatoes to the shopping list", "move the milk to the fridge", "next step" do exactly what the button would, in any language. A message TALC already knows runs straight away as a command, with no AI call; a new phrasing goes to the model, and once it's understood cleanly it's learned as a command too, so next time it's instant and free. Behind this is one principle: the app guides, the model executes. The app decides what happens and checks the result; the model reads what you meant and words TALC's replies. While a table is open, the table itself is the answer, and the chat shows a one-line summary of what changed. Each kind of turn gets its own small prompt and only its own tools, which cut a turn from 14,825 to about 3,000 tokens, so TALC runs on free models.
Built with Next.js, FastAPI, LangChain tool calling, OpenRouter and Groq, hybrid retrieval (embeddings + BM25), SQLite with personal data encrypted at rest.
