Gerard Llanas Logo
● In Development · Open Source

MINCELY

Recipe Intelligence Engine.

Turn any recipe — pasted text, a Word document, a photo, or a YouTube video — into structured data with ingredients, steps, categories and nutrition macros. In seconds.

View on GitHubHow it works
Mincely recipe library
02 / AI Engine

From raw input to structured recipe.

Drop in any source. The AI extracts the title, ingredients, steps, categories and nutrition macros — then surfaces every field for review with inline warnings on anything it's unsure about. You stay in control.

  • Multi-provider routing — Anthropic, OpenAI, Ollama
  • Live editing of generated data before saving
  • Inline warnings on low-confidence fields
  • USDA-backed nutrition with LLM fallback
AnthropicClaude Haiku 4.5 — fast & accurate
OpenAIGPT-4o-mini — broad model coverage
OllamaLocal / offline — no API key required
● Heuristic modeSet `LLM_PROVIDER=none` to fall back to the built-in heuristic parser. Works on plain-text recipes with `Ingredientes` / `Preparación` sections — no API key, no model required.
03 / Import

Any source. One pipeline.

Mincely accepts the messy reality of how recipes actually live: pasted text, screenshots, a Word document your aunt emailed, or a YouTube cooking video. One pipeline, one structured output.

Plain textPaste & parse
.docxWord documents
ImageJPG · PNG · WebP
YouTubeFrom URL
● Vision pipelineImage inputs flow through a vision model (or Tesseract + local LLM in Ollama mode) before structured extraction.
Mincely recipe maker
04 / Edit & Review

You stay in the loop.

Every field generated by the AI is editable before you save. Mincely surfaces low-confidence fields with inline warnings, so you know exactly where to double-check.

Categories and ingredients are auto-detected and chip-tagged — add, remove or rename in one click. Full preview, full edit, full ownership of the data.

Inline warningsLive previewZod-validatedCategories
05 / Recipe Detail

Every recipe, fully readable.

Each saved recipe gets a clean, structured detail page — ingredients, step-by-step preparation, macros, and source metadata. Scroll to walk through the three views.

  • Ingredients with USDA-backed quantities
  • Numbered preparation steps
  • Nutrition macros + per-serving breakdown
  • One-click PDF export of the full page
01 · Header & ingredientsRecipe header and ingredient list
02 · Preparation stepsRecipe preparation steps
03 · Macros & exportRecipe nutrition macros and PDF export
06 / Nutrition

Real macros, from real data.

Mincely queries the USDA FoodData Central API to compute calories, protein, carbs and fat per ingredient — then aggregates per recipe and per serving.

When USDA has no match, the LLM steps in with a structured fallback. You always see which fields came from USDA and which are AI-estimated.

● Quantity parserA custom unit parser converts “2 cups flour” or “300 g chicken” into normalized grams before macro calculation.
● Macros · per servingUSDA + LLM
Calories624 kcal
Protein42 g
Carbs58 g
Fat22 g
07 / Stack

Built modern.

Production-grade architecture from day one — strict typing, schema-first validation, multi-provider AI abstraction, edge-ready database.

Next.js 15
TypeScript
Tailwind v4
shadcn/ui
Framer Motion
Neon Postgres
Anthropic
OpenAI
Ollama
USDA API
Cloudinary
Zod 4
● Status — In Development · MVP shipped

Cook smarter.

Mincely started as a way to organize my father's recipes — it grew into a production-grade AI playground for prompt engineering, multi-provider routing, secure API design and a polished UI. Still evolving, fully open source.

Open Source

MIT-licensed. Fork it, extend it, run it locally.

Roadmap

Auth (Better Auth), Langfuse observability, pgvector recipe search, TikTok / Instagram import.

MVP-only

Anthropic in prod. OpenAI / Ollama wired but not yet deployed.

Open on GitHubAll projects