Recipe apps (and recipe sites, meal kits, cooking channels) serve modern recipes written for today's supermarkets. Food history lives in books, museum exhibits and a few video series you watch but do not cook along with. Lost Menus starts from the real menus people ordered from, tens of thousands of them, and lets you tap any dish and get a cookable recipe rebuilt from its name, with honest notes on what is documented and what is a reconstruction.
The data it runs on
The New York Public Library menu collection: one of the largest historical menu collections, with around 45,000 menus from the 1840s on, many from the collector Frank E. Buttolph. Since 2011 volunteers in the "What's on the menu?" project have transcribed dishes and prices page by page.
Access: NYPL publishes the transcriptions as free bulk CSV files (menus, pages, items, dishes) and the scans through its Digital Collections. We import the CSVs nightly; no scraping.
Limits and gaps: the collection leans heavily on New York and the United States, with ship, hotel and banquet menus over everyday food. Dish spellings vary ("Newburg", "Newberg"), many prices are missing, and menus never include recipes.
Grounding and fallback: recipes are grounded in public domain cookbooks (Ranhofer's The Epicurean, 1894; Farmer's Boston Cooking-School Cook Book, 1896; Escoffier, 1903). Where a dish has no source, the recipe is marked "Reconstructed". Other libraries' digitized menu collections can be added the same way.
1. The mission
Let anyone, anywhere, eat dinner in 1911. Lost Menus turns the world's digitized menus into a kitchen you can actually cook in, one dish at a time, and keeps the food of ordinary restaurants from being forgotten.
2. Why people can't resist it
Time travel you can taste. A real menu, real prices in cents, and tonight you can eat the same thing a diner did on a Sunday in 1911.
The dinner party story. "Everything tonight is from a menu dated April 9, 1911" is the best opening line a host can have.
Collecting decades. Stamps for every decade and city you cook from, and a new menu every Sunday.
3. How it works for real
Menu: NYPL CSVs map each menu page to dishes and prices; scans come from NYPL Digital Collections. A vision model (Gemini Flash or Claude) re-reads any page whose transcription is incomplete.
Dish matching: a fast cheap model (DeepSeek or Gemini Flash via OpenRouter) normalises dish names and clusters spellings across menus.
Recipe rebuild: a frontier LLM (Claude or GPT) writes the recipe, grounded by retrieval over public domain cookbooks, and labels it "Well documented" or "Reconstructed" with the sources it used.
Modern kitchen pass: the same call returns substitutes for hard-to-find ingredients, diet swaps and food-safety temperatures.
Planner: plain code scales recipes, builds the shopping list and the cooking timeline; the price comparison uses the US consumer price index.
Photos: a vision model checks a posted photo is food before it goes on the wall.
4. Cost per use
The menu data is free. One recipe rebuild is a single frontier call of about $0.01 to $0.02, cached for every later cook of that dish, so popular dishes cost nothing after the first time.
Free: the weekly menu, 5 rebuilt dishes a month, the wall. Supper Club ($4 a month): every menu in the archive, unlimited rebuilds, the dinner party planner with printable menus and invitations.
5. Growth loop
Every dinner party sends an invitation card that shows the original menu; guests who eat the food want to host their own. Photos on the menu wall link back to the dish, and each Sunday's menu is a shareable post.
6. Risks and guardrails
Food safety: period methods (raw eggs, low-temperature custards, lard frying) always come with modern safe temperatures and allergy flags.
Honesty: every recipe says whether it is documented or reconstructed. We never claim "the original recipe" without a source.
Attribution: menu scans credit the New York Public Library and the collection; the app follows NYPL's terms for reuse.
Cultural context: some old menus include offensive dish names or practices; those carry a context note rather than silent edits.
7. MVP build order
Import NYPL CSVs into SQLite; render one menu page with tappable dishes.
A Claude prompt that rebuilds a recipe as JSON with sources, confidence and substitutes.
Cache recipes per normalised dish name.
The dinner party planner with scaling and a timeline.
Photo posting and the browse wall by decade and city.