AI Meal Planner: How to Pick One and How to Build Your Own

AI meal planners differ in one way that decides everything: whether the calorie figures come from a food database or are generated by a model. Products look identical from the outside and split cleanly on that question. Here is how to test it in two minutes, and how to build a planner yourself.

How do AI meal planners differ?

Not in the interface, which is why comparison is hard. They differ underneath.

The middle column is the one that matters and the one marketing pages avoid. A product that retrieves figures from a licensed database will say so, because it is the expensive thing they built. Vagueness usually means generation.

The two-minute test

You do not need a review. Three checks tell you which kind you have.

Ask the same food twice. A database returns the same row every time. A generated figure wanders between sessions.

Ask for the source of one number. A database-backed product can name the dataset. An honest assistant will say it came from general knowledge.

Look up an obscure own-brand product, ideally from outside the United States. A real database either has it or says it does not. A model produces a confident figure either way, and that is the failure that matters.

What makes a meal plan actually work?

Separate from data quality, and more predictive of whether you will still be following it next month.

Few enough distinct meals. Three rotating day templates from about twenty ingredients beats seven distinct days. Variety is what people ask for and repetition is what they sustain.

Real constraints, not a diet dropdown. The allergy, the two foods you will not eat, the thirty minutes on a weeknight. This is where chat assistants beat generators decisively.

Gram weights. A plan in servings and bowls cannot be verified or repeated.

Batch structure. Components cooked once and eaten four times is the single highest-adherence pattern, and most planners ignore it entirely because it makes the plan look repetitive on a screen.

Room to be wrong. A plan with no slack fails the first time a dinner runs late. Better ones leave an unplanned meal or a flexible calorie allowance, which sounds sloppy and is the difference between a plan that survives a normal week and one that does not.

A grocery list that shows pack-size rounding. The plan needs 820 g of yoghurt, the shop sells 500 g tubs, so you buy 1,000 g and the spare 180 g either goes to waste or gets eaten. Most tools hide that gap. See building a grocery list from a calorie target.

Does it handle keto, vegan or high protein properly?

Dietary patterns are where the difference between the three kinds shows most clearly.

Chat assistants handle patterns well as arrangement problems. Ask for a vegan week at 150 g of protein and it will reach for the right foods, because the relationships between foods are well documented even when the absolute figures are soft. It will also handle "vegan on weekdays, not at weekends", which no dropdown supports.

Generators handle them as filters. Reliable within the pattern, inflexible at the edges, and the plan is only as good as how the underlying database tags foods.

The failure case for both is a restrictive pattern held for months. A vegan or ketogenic plan optimised for calories and three macros can be quietly short on nutrients that neither the model nor the planner was tracking. If you are eating a restricted pattern long term, check the tool actually carries the nutrients you care about, because a planner that never held a value for iron will not warn you about it. Coverage of that data is discussed in the micronutrient guide.

What to do when it suggests food you will not eat

The most common reason plans get abandoned, and the fix is boring.

Do not ask for a new plan. Ask for a swap that holds the constraint:

Replace the Thursday dinner with something using ingredients already
in the week's list. Keep protein within 5 g and calories within 50.

Reusing the existing ingredient list matters for two reasons. It keeps the grocery list stable, so you are not buying a new item for one meal, and it keeps you inside the set of foods whose figures you already verified.

The broader principle is that a plan should be edited, not regenerated. Regenerating loses everything you checked and produces a fresh set of unverified numbers, which is why people end up with an endless sequence of plausible weeks and no working routine.

Free against paid

Free planners in this category are genuinely usable and most people should start there. A rough plan you follow beats a precise plan you abandon.

What you generally pay for is the food database rather than the planning logic. Generating a sensible week is no longer the expensive part. Branded coverage, provenance, barcode data and keeping all of it current is expensive, which is why products with real food data tend to charge and pure chat wrappers tend not to.

Be sceptical of anything claiming both a large branded catalog and no cost, and look at what it does with your data. Food logs are revealing: eating patterns, household size, often weight history.

A worked week, so the shape is concrete

What a plan that survives contact with a real week looks like.

Three dinner templates, two breakfasts alternating, a lunch that is a batch component plus something fresh. Around twenty ingredients total. Sunday does the cooking: one grain in bulk, one protein roasted, one pot dish for four servings, vegetables prepped. Weeknights become assembly in ten minutes.

The counterintuitive part is that fewer distinct meals produces better adherence. Variety is what people ask for in a survey and repetition is what they sustain in practice. Seven different dinners is seven chances to not feel like cooking. Three, of which two are already made, is far fewer.

Keep one template deliberately lazy, something requiring no thought on the evening you have none, because the alternative to a lazy plan meal is not a better meal, it is a takeaway.

When you want variety, change one template a week rather than rebuilding. The ingredient list stays stable, which means the figures you checked last month are still doing work.

How many calories should it target?

The input everything else depends on, and the one people most often take straight from the same tool that builds the plan.

Get a maintenance estimate with the formula and activity multiplier shown, then apply a deficit or surplus. Published guidance generally puts a moderate deficit around 20 to 25% below maintenance and expresses target rate of loss as a percentage of bodyweight per week rather than a fixed figure, because a 60 kg person and a 110 kg person should not lose at the same rate.

Whatever number appears is a population estimate. Two people with identical inputs differ by several hundred calories a day. Hold it two weeks, track a weekly average weight rather than daily readings, and correct from what the scale did.

Building an immaculate meal plan against a target that is 300 kcal wrong is precision aimed in the wrong direction, and it is the most common way this whole exercise fails.

Building your own

The third route, and it suits people who want the conversational planning without the invented numbers.

Use a chat assistant for the arrangement, which is the thing it is genuinely best at, and attach a food database so the figures stop being generated. The assistant searches the catalog, gets a row with macros per 100 g, and scales it to your gram weight.

That is this nutrition MCP server. Stated plainly: it uses an API key header and is verified with Claude Code and Cursor, not inside ChatGPT, since browser sign-in is not enabled here. Tools and arguments are in the docs and the plan is on MCP pricing. The full workflow is in meal planning with a real nutrition database, and the recipe side is in recipe macros without a spreadsheet.

If you are building a product rather than planning your own meals, you want HTTP endpoints, which is a REST plan.

Where all of them are weak

Adherence. None of them makes you cook the food.

Your actual energy needs. Every one starts from a formula estimate, and two people with identical inputs differ by several hundred calories a day. Whatever target it uses is a hypothesis to correct over a fortnight.

Micronutrients. Most optimise calories and three macros. Fibre, sodium and iron are often absent or sparse, and a plan reporting zero for a nutrient it never held is worse than one that says it does not know.

Clinical situations. A calorie arrangement is not a therapeutic diet. With coeliac disease, kidney disease, insulin-dosed diabetes, a diagnosed allergy, an eating-disorder history, or in pregnancy, a qualified human needs to review any generated plan. A stated allergy is a preference filter, not a safety system.

Cost. None of them knows what food costs where you shop. Price data is local, seasonal and changes weekly, so a request for a cheap week produces guesses. What they do handle well is a swap where you supply the constraint, since protein density and similar relationships hold up even when absolute figures are soft.

Food you have already. Almost none of them start from your cupboard. A plan that ignores the half bag of lentils and the vegetables about to turn generates waste, and waste is the quiet cost of meal planning that nobody puts in the comparison tables.

Households. Most of these plan for one person. Cooking for a family with different targets is the normal case and the badly served one, and the practical answer is usually a shared base with portions adjusted rather than separate plans.

What we have not measured

We have not tested the meal planner products in this category side by side, measured database coverage against a fixed basket, or compared plans against weighed food. No rankings appear here because we have not earned them.

The two-minute test above works on any product including ours, which is more useful than a score that would be stale within a quarter. When we do run a coverage measurement it will be published with the basket and the method, so you can re-run it rather than trust it.

Related

• AI diet plan generators: what they actually do
• Meal planning with a real nutrition database
• ChatGPT meal plan: prompts, meal prep and grocery lists
• A weekly grocery list built from a calorie target