1. Hard to plan
Hitting 1,800 calories and 120 g of protein means adding up every label. AI chatbots just guess.
“Where did 420 calories come from?” The chatbot made it up.
Product story
Why this exists and how it’s built
The problem, who it’s for, where it sits among existing tools, how it works in five levels, what building it taught me, what it will never do, and what’s still missing.
Summary
An hour of chores, every week
Every week, people who eat to a goal spend an hour or so on the same chores: adding up calories from labels, checking what’s in the fridge, writing a grocery list, and searching a store’s website for each item.
CartMyMeals plans your week around your goals, skips what you already have, and shows how every number was worked out. Then it fills your own Amazon or Walmart cart.
Why and how I built this
Why
Eating to a goal is a weekly chore.
I wanted one path from goal to cart.
How
- Built with ClaudeAs my coding partner, and in the product where it helps
- Sourced nutrition dataEvery ingredient traced to USDA or a standard table
- GuardrailsIt never checks out; you pay
- 6 evals test itDiets, allergies, nutrition, pantry, cart, AI weeks
01
The problem
Planning meals to a goal is hard. Even a good plan fails without the ingredients.
Eggs
Spinach
Rice2. No ingredients
Some of it is at home. The rest waits for a store run, or an order you put off or forget.
“I have the plan, but not the ingredients until Saturday.”
3. Goal missed
So you swap in what you have, like pasta for chicken, and the week drifts off target.
“I swapped in pasta and fell 30 g short on protein.”
02
Who it is for
Three people I designed for, each mapped to a feature that serves them. Hypotheses Being checked in a private beta of about 150 people.
MayaThe macro tracker · Trains four times a week
Goal
1,800 calories and 120 g protein a day, without a spreadsheet.
“I spend Sunday adding up labels, and I still don’t trust the numbers.”
What helps
Every number computed from USDA data, with the math shown.
Priya and SamThe family planners · Two kids, one with a nut allergy
Goal
One menu the whole family eats, sized for the kids.
“By the time I’ve checked every label and built the cart, dinner is late.”
What helps
Kids’ portions and allergies, and the whole list in one Walmart cart.
JordanThe plant-based cook · Vegan, with a full pantry
Goal
Enough protein, without wasting what’s already at home.
“Half my lentils expire because I forget I bought them.”
What helps
Vegan filters, and a week planned around the pantry first.
Willingness to pay Hypothesis Strong interest in anything that saves one to two hours a week and removes nutritional guesswork. Not yet tested.
03
The landscape
ManualAutomated →↑ Verifiable numbersMade-up numbers
MyFitnessPal, CronometerAccurate, but logged by hand after the fact
Mealime, HelloFreshFixed recipes, no pantry arithmetic
ChatGPT, Claude promptsFast ideas, recalled numbers, no store
MyFitnessPal, CronometerAccurate, but logged by hand after the factMealime, HelloFreshFixed recipes, no pantry arithmeticChatGPT, Claude promptsFast ideas, recalled numbers, no storeCCartMyMealsComputed nutrition, pantry-aware, a real cart
Trackers get the numbers right after you’ve eaten. Chatbots plan ahead but recall their numbers. Meal kits ignore what’s in your fridge. No one plans ahead with verifiable numbers and ends at a cart.
04
The solution, in levels
Each level removes one kind of risk before the next is added. Levels 1 to 3 are built. Levels 2 and 3 run on the try-it page and through the MCP server, so the same engine works inside Claude.
- 1
LLM answerer (the baseline)Benchmark
Profile in, seven-day plan out. The model invents meals and recalls their calories. Kept deliberately, as a live benchmark: its figures disagree with Level 2’s by roughly 15% on calories and 17% on protein.
- 2
Retrieval with computed nutritionLive
The model may only pick recipe ids from a shortlist. Every calorie is arithmetic in a database view: 60 g oats × 380 kcal/100 g = 228 kcal. Live for visitors: start with one sentence and the week is arranged to your wishes, then change it in your own words (“swap beef dinners for tofu”). Claude only picks from meals the rules allow, and any id outside them is dropped.
- 3
Pantry shortfall and a scoped cartLive
Need minus have, in grams, across three unit systems (grams, packages, shoppable units). Each gap is matched to a real product by a scorer; weak matches are declined and shown, never guessed. The cart is built and handed over at checkout, on the demo store or as one link that fills the shopper’s own Walmart cart. The same tools are offered to AI apps as an MCP server, so a week can be planned and shopped from inside Claude (cartmymeals.com/mcp).
- 4
OrchestrationStarted
Claude already has five jobs, each with a model chosen by comparison: Sonnet reads a household sentence and receipts, arranges the week and turns requests into changes; Haiku lists a dish’s groceries; rules check each one. Still to build: one controller over them, the two approval gates as real steps, and a visible run trace of which model ran, what it cost and where it waited for a person.
- 5
Preference rerankerNot started
A small model trained on logged Keep / Swap / Skip decisions, reranking candidates without loosening the guardrails. Waiting on enough real decisions; they are logged from day one and cannot be backfilled.
How it’s built
1 · Inputs
Profileone sentence that Claude turns into the profile, the pantry and this week’s wishes, or the form; the result is shown to check
Pantry and own dishestyped rows or a receipt photo read by Claude; any dish name, with Claude suggesting its groceries to check
2 · Plan
Postgres (Neon)ingredients pinned to USDA records; recipes as gram weights; nutrition is a SQL view
Plannerretrieval and ranking with diet, allergy and carb rules, plus anything left out this week. From a wish or a request in words, Claude Sonnet arranges or changes it using only these options (Level 2, RAG)
3 · Shortfall
Need minus havein grams, scaled for the household, rounded to what stores sell (eggs by the dozen)
Product matchinga scorer picks a product or holds the item back, never guesses
4 · Cart
Amazon Livethe whole list in one click, matched by Amazon Fresh or Whole Foods
Walmart Livea real store: one add-to-cart link fills the shopper’s own cart
Shopify Storefront APIsearches about 630 products copied from Walmart, to match each item and price the list
MCP server Livethe same plan, swap and shopping-list tools for Claude and other AI apps; no payment tool
Kroger, Instacart, DoorDashtest cart, or item-by-item links
Guardrails:No checkout tool: a person always paysYou approve the menu firstThe model picks recipes, never numbersLow-confidence matches go back to youAllergies filter every mealAll nine →Evals: five automated checks run against the real planner and storeDiet rules 80/80Allergies 20/20Nutrition fit 35/36Pantry reading 42/42Cart matching 92/92How I test it →
Next.js 16 and React 19 on Vercel · TypeScript · Postgres on Neon · Walmart prices via SerpApi · Shopify Storefront GraphQL · Claude Haiku 4.5 and Sonnet 5.5 · MCP (Streamable HTTP) · USDA FoodData Central
Integrations
8 grocery integrations built or evaluated. Two are live for shoppers, who pick the store: Amazon (through Amazon Fresh or Whole Foods) or Walmart, each taking the whole list; ordering sits behind one interface (a shortfall list in, a cart out), so the rest can drop in without touching the planner.
AmazonOne click sends the whole list to Amazon’s ingredients page for Amazon Fresh or Whole Foods, which matches each item for the shopper to review. A regular Amazon cart link remains for items Amazon sells itself. Try it →Live
WalmartOne add-to-cart link puts the whole list in the shopper’s own Walmart cart. Try it →Live
CartMyMeals MCP serverThe product as tools for AI apps: plan a week, swap meals and get a Walmart cart link from inside Claude, with the same rules and no payment tool. See /mcp. Try it →Live
Shopify Storefront APIFixed queries match and price every item against about 630 products copied from Walmart. Used for every visitor.Working- Shopify UCP over MCPAn AI agent discovers the store’s tools and builds the cart itself. Owner-only, because the tools include checkout.Owner-only
KrogerOAuth with customer consent; writes to a cart in Kroger’s test environment only.Test only
InstacartReturns a list link with no ordering endpoint; the developer program is closed to new apps.Evaluated
DoorDashIts only way to order is a command-line tool (dd-cli) in invite-only beta; there’s no public cart API yet.Waiting on beta
TargetNo add-to-cart link for a whole list, so it would be item-by-item links at best.Evaluated
The public version
Anyone can try it, with no account:
- Describe your household in one sentence, or fill in the form. CartMyMeals plans the week from it, and you can change anything in your own words.
- List the pantry, or skip it. The week is planned around what’s at home.
- Edit and approve. Replace or remove meals; nothing is listed until you approve.
- Fill your Amazon or Walmart cart, or save the list for any store.
By default the week is chosen by retrieval and rules, with no model call, so it’s instant. Ask CartMyMeals to arrange or change it and it picks only from the recipes the rules retrieved; every number is still computed. It also reads what people write or photograph, into choices they check.
Example of what it produces
05
What building it taught me
Six things I’d take into the next product.
- ~15%AI guesses numbersA chatbot’s calories were about 15% off real food data.
- 7 of 20Tests catch what you missTyping “soy” didn’t remove tofu until a test caught it.
- 2 mix-upsWords are tricky“Almond milk” was read as almonds.
- 99 of 99Show your workEvery calorie traces back to a cited food record, so a number can be checked, not just trusted.
- 0 wrong“Not sure” beats a guessWhen the matcher isn’t confident, it hands the choice back. In testing, not one wrong product reached a cart.
- 6 changesTesters find what tests don’tEvery eval passed, yet testers said the meals repeated and didn’t fit their families. Their notes became six changes in a week.
06
What testers told me, and what changed
A private beta since September 2026. Each change below shipped within days of the feedback.
~150visitors
43planned a week
16sent a real cart
10wrote back
The meals repeat way too often. I have the same lunch almost every day.
Beta tester✓ What changedMore variety
Five breakfasts, six lunches and a different dinner each night, and a note when choices are too narrow to avoid repeats.
I would never eat chicken breast at breakfast.
Beta tester✓ What changedBreakfast that makes sense
No meat or fish at breakfast by default, and a choice of sweet or savory.
I would never buy Atlantic salmon, only wild caught.
Beta tester✓ What changedShopping preferences
Organic, grass-fed, free-range and wild-caught swap the products that go in the cart.
The meals were very different from what my family eats.
Beta tester✓ What changedTheir food, not ours
Cuisines they love, meats they eat (no beef, no pork), and 54 more recipes, from dal makhani to chicken tinga tacos. Indian dishes appear only for those who pick Indian.
Ask for favorites.
Beta tester✓ What changedFavorites and own dishes
Favorite dishes repeat every week. Any dish, recipe link or YouTube video can go anywhere in the week, with its ingredients in the cart.
Items added to my Walmart cart were not the correct amounts: 12 broccoli crowns.
Beta tester✓ What changedStore-sized amounts
Quantities are rounded to what stores sell, and anything not found is listed so nothing goes missing quietly.
What the numbers said. Most people who finish the first screen plan a week, and about one in three of those send a cart. The first screen lost the most people, so it got shorter: optional preferences fold away, the week reads one day at a time on a phone, and people can now start with one sentence instead of a form.
07
Guardrails: What the agent cannot do
The principle is safety by architecture, not by prompt.
No purchasing
Checkout tools such as complete_checkout are never wired in. The agent can stage a cart; only a person can pay.
Show what was sent
Some cart APIs are write-only and can’t be read back, so every payload is recorded and shown, and test environments are labeled as test.
No silent substitutions
Below a confidence threshold the match is declined, the reason is shown, and the choice goes back to the person.
You approve first
Nothing goes on a shopping list until you approve the week. Whatever CartMyMeals reads or changes (a sentence, a dish, a receipt, a request) comes back as choices you can see and undo, and its use is capped per visitor per day.
Allergies cover the whole table
“Soy” removes tofu, edamame and soy sauce, not just the word. A child’s allergy applies to everyone’s meals, since the family eats together. The allergy eval checks this: 20 of 20.
Says no instead of guessing
If no meal fits your diet and allergies, it tells you rather than bending a rule. A pantry line it can’t read, or a “box” with no known weight, is flagged, not guessed.
The model can’t invent numbers
It picks recipe ids from a retrieved shortlist and never writes a calorie. An id outside that list is dropped and reported; every figure is computed from gram weights.
Edits can’t bend the rules
Replace and Add only offer meals that already passed your diet and allergy filters, and the server re-checks every choice, so an edited request can’t bring a meal back in.
Warns before it adds
Walmart’s link adds to whatever is already in your cart and can’t empty it, so the page tells you to empty it first rather than letting old items slip through.
08
Success criteria
These are targets, not measured results. The app has not yet run at a scale that could measure them.
| Metric | Target | Where it stands |
|---|---|---|
| Weekly planning and ordering time | From ~60 minutes to under 10 | Not yet measured |
| Nutrition figures produced by a model | Zero | Met by design Every figure is computed from gram weights, using USDA values for 51 of 99 ingredients and standard food-composition tables for the rest. |
| Product match precision | Over 90% of matches accepted first time; no unflagged wrong substitutes | Partly built In the offline eval every stocked ingredient matched the right product and none matched a wrong one; acceptance by real shoppers not yet measured |
| From list to a real retailer’s cart | One click, no retyping | Met for Amazon and Walmart The whole list goes to the shopper’s cart in one click: Walmart’s add-to-cart link, or Amazon Fresh and Whole Foods through Amazon’s ingredients page. Instacart still goes item by item |
| Food waste | Less leftover food each week, by netting out what is on hand | Partly built Pantry netting built; waste not yet measured |
09
Evals: How I test it
Six evals run the same code the site runs and judge it against rules written out separately, so a bug in the planner can’t also hide itself from the test. Five call no model, so they’re free and give the same answer twice. The sixth has Claude arrange weeks for three households and checks every pick came from the rule-filtered shortlist and still respects the diet and allergies. Last run October 6, 2026.
| Eval | What it asks | Result |
|---|---|---|
| Diet and style rules safety check | Across every diet, style, gluten-free and protein combination, does any meal or Replace option break the rules? | 80 of 80 |
| Allergies and dislikes safety check | When someone types an allergen in "Avoid", does any meal or Replace option still contain it? | 20 of 20 |
| Nutrition fit | Does the week land within 15% of the calorie target, and within 10% of the protein goal when one is set? | 35 of 36 |
| Pantry reading | Does typed pantry text become the right ingredient and amount, and is anything unknown left unrecognised rather than guessed? 0 non-food or unknown items were wrongly matched | 42 of 42 |
| Cart matching safety check | For every ingredient the store stocks, is the right product put in the cart? Is anything put in the cart that is wrong? 0 wrong products in the cart, 0 held back for a person to choose, 5 not stocked | 92 of 92 |
Findings so far
<1 minfrom Try it now to a cart
92/92right store product, none wrong
269recipes, nutrition computed from grams
1 clickwhole list into a Walmart cart
Since Oct 1, visitors built 115 carts and sent 10 to Amazon and 12 to Walmart. These measure the product. A private beta with real shoppers is running now: see what they told me below.
Bugs the tests caught
ProblemSomeone avoiding soy still got tofu. Avoiding milk still gave yogurt.
FixedAn allergy now covers every food made from it.
Problem“Almond milk” in the pantry was read as almonds.
FixedThe pantry reader tells them apart.
Problem“2 cans black beans” was read as dried beans.
FixedCanned beans count as cooked beans.
ProblemA vegetarian week included a soup made with chicken broth.
FixedBroth now counts as meat.
ProblemA gluten-free week included bread crumbs.
FixedBread crumbs now count as gluten.
ProblemThe shopping list showed a different tortilla than the one added to the cart.
FixedThe list now shows exactly what goes in the cart.
ProblemA cheaper model looked as good at reading receipts, but read 10 oz of spinach as 10 g.
FixedThe model test now checks amounts, and receipts stay on Sonnet.
Still not perfect: 1 of 36 nutrition checks fails. A 1,400-calorie, high-protein vegan week comes out about 15 g short on protein. The test stays strict, so the score only goes up when the recipes get better.
Which model for which job
The same test inputs went to three Claude models for each AI task: accuracy, then seconds and cost per call. The rule: the cheapest model that scores as well as the best. Bigger wasn’t better on most tasks.
| Task | Haiku 4.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|---|
| Describe household | 98%1.2 s · 0.17¢ | 98%2.2 s · 0.39¢In use | 100%2.5 s · 0.81¢ |
| Suggest ingredients | 90%1.2 s · 0.13¢In use | 93%1.9 s · 0.29¢ | 93%2.2 s · 0.59¢ |
| Read receipts | 92%2.4 s · 0.14¢ | 100%2.4 s · 0.34¢In use | 100%3.5 s · 0.91¢ |
| Arrange week | 100%6.3 s · 0.59¢ | 98%8.4 s · 1.80¢In use | 93%11.8 s · 3.93¢ |
Small samples (10 sentences, 10 dishes, 5 receipts, 3 weeks), and the week test checks safety and variety, not taste. The first version of this test only checked that receipt items were found, and Haiku looked as good as Opus. Checking amounts too showed Haiku reading ounces as grams (10 oz of spinach as 10 g), so receipts stay on Sonnet. Haiku lists groceries, where it matched the larger models; Sonnet arranges the week until a test of week quality, not just safety and variety, says otherwise.
10
Limitations
What it can’t do yet, stated plainly.
Food data4
- Not every value is from USDA51 of 99 ingredients use USDA values; the rest use standard food-composition tables. Protein bars are branded, and cooking losses aren’t modelled.
- Strict diets still rotate269 recipes now, 18 with fish, but a vegan keto week still has only a couple that fit. The planner says so rather than padding the week.Roadmap: More recipes for narrow diets
- Portions scale, within limitsEach recipe scales from 0.75× to 1.5×, so weeks land close to the calorie goal. Low-calorie, high-protein vegan weeks still fall 10–15 g short on protein.Roadmap: More recipes for narrow diets
- The pantry reader knows only its own ingredientsUnknown items are ignored, and a “box” with no known weight is flagged.Roadmap: Pantry from a photo
Store and safety3
- Walmart’s cart is a one-way linkThe link adds to your Walmart cart but can’t empty it first or read it back. Items already there stay, and fresh food needs a local store set.Roadmap: A real store, confirmed
- Allergies are matched by nameIt can’t see inside a packaged protein bar or wrap.
- Amazon picks the productsAmazon Fresh and Whole Foods match each item themselves, so pack sizes can be larger than needed, and items not stocked locally are flagged for the shopper to swap. A browser opens one tab per click, so one cart is the whole list.
Proof2
- The AI suggests, rules decide what’s allowedCartMyMeals reads a household sentence, a dish or a receipt, arranges the week and changes it on request, but only with meals the rules allow. Arranging adds about ten seconds; without a wish, the instant rule-based week is used.Roadmap: Level 4, orchestration
- A small private betaAbout 150 people since September 2026. Feedback is qualitative; time saved isn’t measured yet.Roadmap: Measure time to cart
11
Roadmap
- Shipped
- ✓Levels 1–3Model baseline, computed nutrition, pantry shortfall and a scoped cart.
- ✓Amazon in one clickThe whole list to Amazon Fresh or Whole Foods, matched by Amazon.
- ✓Walmart one-click cartThe whole list into a real Walmart cart.
- ✓Sourced nutrition dataAll 99 ingredients traced to USDA or a standard table.
- ✓6 evalsDiets, allergies, nutrition, pantry, cart and AI-arranged weeks, re-run with every change.
- ✓269 recipes, scaled portionsMyPlate Kitchen recipes, dishes testers asked for, and a full week for every cuisine and diet; portions scale to the target.
- ✓Taste and favoritesCuisines, meats, shopping preferences, favorite and own dishes.
- ✓A phone-first weekOne day at a time, with a tap menu for each meal.
- ✓An MCP serverPlan, swap and shop from inside Claude or any MCP app, with the same guardrails.
- ✓Talk to itStart with one sentence, change the week in your own words, get any dish’s groceries, read a receipt; it only uses meals the rules allow (RAG).
- NowI am here
- Private betaRunning since September 2026, about 150 people so far. Next: measure time to cart and matches accepted.
- Re-measure Level 1Re-run its ~15% gap against the USDA-based figures.
- Next
- More recipes for narrow dietsHigh-protein vegan and keto, so strict weeks stop rotating through a handful.
- A real store, confirmedA partner API that reads the cart back and offers swaps.
- Pantry from a photoSnap the fridge; a vision model fills the pantry to check.
- A weekly budgetSet a spend limit in your profile; the planner favors cheaper meals and shows the list against it.
- No repeatsAn option for a week where no meal repeats. Today each slot rotates five to seven recipes.
- Later
- Level 4, orchestrationSeparate models for ideas, rules and cart, with a visible trace.
- Level 5, preference rerankerLearn from Keep / Swap / Skip without loosening a guardrail.