Skip to content

Connector documentation

Connect an AI assistant to Live Better AI

Assistants can search our recipes, suggest a week of meals and make an Instacart shopping list, with no account at all. With a member's permission they can also read and change that member's own week.

Endpoint
https://www.livebetterai.com/api/mcp
Transport
Model Context Protocol over Streamable HTTP, stateless. Protocol version 2026-07-28, and 2025-era clients through initialize.
Browsers
Call from a server, or from a browser page on one of these origins, which get CORS answers (preflight included): https://www.meta.ai, https://meta.ai, https://www.muse.ai, https://muse.ai, https://platform.muse.ai. Every other browser origin is refused with HTTP 403. From those pages the tools that need no account work, with or without a partner key. Member sign-in needs a server-side client: /.well-known/oauth-protected-resource sends no CORS headers.
Where
The United States only.
Price
Live Better AI is free: no subscription, no trial and no payment to us; groceries ordered from its Instacart shopping lists, and any fees, are paid at the Instacart checkout.

Using Live Better AI in Claude

Live Better AI finds recipes, suggests a week of meals and makes Instacart shopping lists right in your Claude chat. Live Better AI is free, with no subscription, no trial and no card.

Add it

  1. In Claude, open Customize, then Connectors, and add Live Better AI from the directory.
  2. If Live Better AI is not in the directory yet, choose Add custom connector and paste https://www.livebetterai.com/api/mcp. Recipes, weeks and shopping lists work that way without signing in.
  3. In a chat, press +, then Connectors, and check that Live Better AI is turned on.
  4. Name Live Better AI in your message so Claude knows to use it.

Without signing in

Ask for recipes and Claude shows them with photos, cook time and nutrition. It can open a recipe's ingredients, suggest a week of meals for your household and swap any meal you do not want, and make an Instacart shopping list for the meals you pick. You tap the link, look over the list on Live Better AI, then press Shop ingredients and check out on the Instacart page yourself. Nothing is saved to an account, and what you ask is not tied to you. A recipe's steps need your Live Better AI account.

With your Live Better AI account

  • The first time you ask for your own week or grocery list, Claude shows a Connect button. Press it and sign in on livebetterai.com with your Live Better AI account.
  • Read the approval page, which names the app and what it can see and do. Choose Full access or Read only, then Allow access. Claude then carries on with what you asked. You can switch later on livebetterai.com, in Settings, then Connected apps.
  • Once connected, Claude can show your week, any meal in full with its steps, and your grocery list. With Full access it can also swap a meal (it shows you choices and swaps the one you pick, and your grocery list updates) and make an Instacart list for your week. Read only lets it look at your week, meals and grocery list, but not swap meals or make Instacart lists.

Things to ask

Each of these names Live Better AI, so Claude knows to use it.

  • Use Live Better AI to find high-protein chicken dinners in 40 minutes or less.
  • What do I need to make the Enchilada Chicken from Live Better AI?
  • Ask Live Better AI for five vegetarian dinners for a family of four.
  • Make a Live Better AI Instacart list for the first three of those dinners.
  • Show me my Live Better AI meal plan for this week.
  • What's on my Live Better AI grocery list?
  • Ask Live Better AI to swap Wednesday's dinner in my week.

If something goes wrong

Nothing happens, or Claude answers without Live Better AI.
Name Live Better AI in your message, for example "Use Live Better AI to find a quick dinner." Then check it is turned on for the chat: press +, then Connectors.
The sign-in window does not open, or sign-in does not finish.
Your browser may have blocked the window. If you use Claude in a browser, allow pop-ups for claude.ai. Press Connect again and sign in the way you joined Live Better AI: your email and password, or Google. If you joined with Apple and the page has no Apple button, choose Forgot password? to set a password first.
Your week is empty.
Claude can only show a week you have built. Build a week in the Live Better AI app first, then ask again.
You want to disconnect.
Remove Live Better AI from your connectors in Claude (Customize, then Connectors), and disconnect it in Live Better AI too: Settings, then Connected apps.

Claude Code cannot sign in to your Live Better AI account yet. It can still find recipes, suggest a week and make Instacart lists without signing in. To see your own week and grocery list, use Claude on the web, the Claude desktop app or the Claude phone app.

Still stuck? Email support@livebetterai.com.

Claude is a trademark of Anthropic, PBC.

Two levels

No account

No sign-in and no credential. Four tools: search the catalog, read a recipe, suggest a week and make an Instacart shopping list. Nothing about the person is stored. A recipe’s method is on its recipe page for Live Better AI members (a free account: sign in on the recipe page to read it). Partner keys and signed-in members get it in the answer.

Account

The member signs in on livebetterai.com and approves the app. The assistant then also reads that member's week, meals and grocery list, swaps a meal and makes an Instacart list for the week. Every tool tells the assistant to ask the member before a change; whether it does depends on the assistant and its settings, since some let a person allow ordinary changes without asking each time. Making an Instacart list is a Sensitive write. When approving the app, the member chooses Full access or Read only (see Read only below).

Read only

When a member approves an app on our consent page, they choose Full access or Read only, and they can switch at any time in Settings, then Connected apps, on livebetterai.com. We store the choice for that member and app and apply it on every request. Nothing the app sends changes it.

  • Under Read only, tools/list leaves out every Write and Sensitive write tool: shop_recipes, set_up_household, build_my_week, edit_my_pantry, swap_my_meal, swap_meal, edit_my_grocery_list, add_favorite, import_recipe, create_shopping_list. Every Read tool stays.
  • A call to one of them, by any of its names and inside a batch too, runs nothing and answers read_only in the tool answer. Tell the member they can switch the app to Full access.
  • A switch applies from the app's next tool call, whatever token it holds. A tools list fetched before it may be up to a minute old. The server sends no notifications/tools/list_changed, so after a switch to Full access, fetch tools/list again (a new chat).
  • The no-account level has no Read only choice: it changes no one's account.

Authentication

  • None. The no-account tools answer a request with no credential.
  • Partner key, on request. We issue a key to an assistant maker. Send it in any one of these headers: X-Api-Key: lbai_pk_..., Api-Key: lbai_pk_..., Authorization: Bearer lbai_pk_..., Authorization: ApiKey lbai_pk_..., Authorization: lbai_pk_... (no scheme). It gives the no-account tools their own budget and recipe methods. It is not a sign-in: account tools still need the member, and a member's token in Authorization: Bearer always wins over a key header. A value in X-Api-Key or Api-Key that is not one of our keys is ignored.
  • OAuth 2.1, for a member's account. Authorization code with PKCE (S256), and refresh tokens. Clients are registered with us in advance; dynamic registration is off. The member approves the app on our consent page. Access tokens last an hour. Disconnecting in Live Better AI ends access within a minute, and a connection unused for 90 days is disconnected.

Discovery

Protected resource (RFC 9728)
https://www.livebetterai.com/.well-known/oauth-protected-resource/api/mcp
Authorization server issuer
https://auth.livebetterai.com/auth/v1
Server metadata (RFC 8414)
https://auth.livebetterai.com/.well-known/oauth-authorization-server/auth/v1

Calling an account tool without the member's sign-in answers HTTP 401 with this challenge, which is how a client knows to start the sign-in:

WWW-Authenticate: Bearer resource_metadata="https://www.livebetterai.com/.well-known/oauth-protected-resource/api/mcp"

Without a credential, tools/list also lists these account tools, so a client that calls only listed tools can reach that challenge: get_weekly_plan, get_meal_details, get_grocery_list, swap_meal, create_shopping_list. With a partner key it lists the no-account tools only. A single tools/call that carries OpenAI's request metadata (_meta keys starting openai/) gets the same challenge in the tool answer instead: an unauthorized error with _meta["mcp/www_authenticate"], which is where OpenAI's clients start the sign-in.

Tools

Read changes nothing. Write changes the member's Live Better AI data. Sensitive write makes something outside Live Better AI: an Instacart shopping-list link. Every answer is { ok: true, data } or { ok: false, error }, as structured content and as text. Muse is offered the tools marked in the last column.

Muse is a trademark of Meta. Our Muse directory listing is in review.

ToolLevelLabelWhat it doesMuse
search_recipesNo accountReadSearch the Live Better AI recipe catalog, screened for stated allergies and diet (and a signed-in member's own, by default).Yes
get_recipeNo accountReadOne catalog recipe: photo, time, servings, nutrition and ingredients; the method for partners and signed-in members.Yes
suggest_weekNo accountReadA varied week of catalog recipes for a household size, with nothing stored.Yes
shop_recipesNo accountSensitive writeCreate an Instacart shopping-list link for catalog recipes. No order or payment is placed. US only.Yes
set_up_householdAccountWriteReview or save the member's household setup.No
build_my_weekAccountWriteBuild a new week, or rebuild the rest of one.No
get_week_build_statusAccountReadCheck a week build.No
get_my_pantryAccountReadRead the member's pantry.No
edit_my_pantryAccountWriteAdd, remove or re-date pantry items.No
get_my_planAccountReadRead the member's week (this week by default).No
get_weekly_planAccountReadThe same tool as get_my_plan. Read the member's week (this week by default).Yes
get_meal_detailsAccountReadOne planned meal in full, by plan, day and meal.Yes
swap_my_mealAccountWriteSuggest, then swap in, a recipe the app vetted for that meal.No
swap_mealAccountWriteThe same tool as swap_my_meal. Suggest, then swap in, a recipe the app vetted for that meal.Yes
get_my_grocery_listAccountReadRead a week's grocery rows as an edit names them.No
get_grocery_listAccountReadRead the week's grocery list as two lists, to buy and may already have, a page at a time.Yes
edit_my_grocery_listAccountWriteCheck, exclude or adjust grocery rows.No
list_my_favoritesAccountReadRead the member's saved recipes.No
add_favoriteAccountWriteSave a recipe as a favorite.No
import_recipeAccountWritePreview, then save, a recipe from a web page.No
create_shopping_listAccountSensitive writePreview, then create, an Instacart shopping-list link for the week or chosen meals. No order or payment is placed. US only.Yes

search_recipes

Read · No account

Searches the Live Better AI recipe catalog, best match first unless another sort is asked for. Changes nothing. With a member's sign-in, that member's own allergies and diet apply too, and anything stated can only narrow them.

Inputs

  • query: What the person asked for, in their own words, up to 120 characters (for example "high-protein chicken dinner" or "kid friendly dinners"). Words naming a meal, a diet, a time, a protein or calorie wish, a family or a household ("for two") are read as filters or an order, and understood in the answer says how; the other words are matched against titles, ingredients and cuisines, and any one of them is enough. A preparation word such as ground or fresh never matches on its own: "ground beef" matches beef, recipes that name ground beef first. Give a query, a diet, a meal_type, a cuisine, or a protein, calorie or time limit.
  • diet: vegetarian, vegan or pescatarian. Checked against ingredient names and the recipe's verified diet flags.
  • allergies: Up to 10 of milk, egg, fish, shellfish, tree_nut, peanut, wheat, gluten, soy, sesame. Whenever the person mentions any allergy, intolerance, sensitivity or food they must avoid, pass each allergen here on every call, even if a query names it too (any other food goes in exclude_ingredients): reading a query is only a fallback. Matched on ingredient names only.
  • exclude_ingredients: Up to 10 other foods to keep out, as plain words (for example "cilantro").
  • max_minutes: 5 to 240, by our recipe audit's time estimate. A recipe with no audit time never matches.
  • cuisine: One of the catalog's cuisines (for example italian, mexican, thai, indian, mediterranean). Only recipes the catalog tags with it; an untagged recipe does not match.
  • min_protein_g: 1 to 200. At least this many grams of protein per serving, as the app rounds it. Unknown protein does not match.
  • max_calories: 50 to 3,000. At most this many calories per serving. Unknown calories do not match.
  • meal_type: breakfast, lunch (salads, soups, sandwiches and other lunch dishes), dinner (main dishes only, never a side), snack (never a main), dessert or side.
  • sort: relevance (the default: the best match first, then a sane cook time), catalog (the catalog's own order; results match the search but are not ranked by it), quickest, most_protein or fewest_calories. Every sort but catalog orders the whole match, not one page; recipes with no number for the sort come last, and ties are broken the same way every time, so the order is stable.
  • limit: 1 to 20 results. Default 10.
  • cursor: next_cursor from the previous page. It works only with the search and sort it came from.
  • servings: How many people the person is cooking for, 1 to 12. It never changes which recipes match. In a chat that draws cards, every recipe the card opens and every list it makes is for that many. In one that does not, the answer is sized for it and says for_servings; left out there, a household the query names ("for a family of four") is used.

Returns

recipes (summary cards: recipe_id, title, image_url, recipe_url, cook_time_minutes (our recipe audit's estimate), calories_per_serving, protein_g, carbs_g and fat_g (the panel the app shows, in whole numbers), servings and cuisines, and diet_labels; no ingredients or steps), next_cursor, sort, understood (the words matched, the meal_type, diet labels, carb cap and order the query asked for, and one plain line per phrase it read), filters (the cuisine, protein and calorie filters applied), screened_for (what was screened, never who asked), screen_basis "ingredient_names" when a screen applied, for_servings in a chat that draws no card when the search was sized for a number of people, and allergy_note ("Allergy screens match ingredient names only, so check labels.") when an allergy, food or diet screen applied, including one the signed-in member saved. The answer's second text block depends on the chat. In a chat that draws cards (a member signed in on any app but Muse's own, or, with no credential and no Meta origin, a client that declares the MCP Apps UI extension, io.modelcontextprotocol/ui, or that the server knows draws cards), it is a note for the assistant with each recipe's ingredient names, and the card shows the photos. In a chat that does not, and on a partner key or a page on meta.ai or muse.ai, it is written for the person, to pass on as it is: "Sized for 4 people." when it was, then one short block per recipe (its markdown photo line first, then its title, time, calories, protein and servings, then its recipe page link), and the allergy line last when the answer carries it. No ingredient list, and a number we do not have is left out, never shown as 0. Signed in on Muse's own app, it has one markdown image line per photo.

Errors

invalid_input, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "search_recipes",
  "arguments": {
    "query": "high-protein chicken dinner",
    "allergies": [
      "peanut"
    ],
    "max_calories": 600,
    "limit": 5
  }
}

get_recipe

Read · No account

Reads one catalog recipe. Changes nothing. Without an account the method is withheld, as on the public recipe page; a partner key and a signed-in member get the steps.

Inputs

  • recipe_id (required): Digits only, from search_recipes or suggest_week, as a string or a whole number.
  • servings: How many people to cook for, 1 to 12. Every amount comes back scaled from the recipe's own yield, as the app scales a recipe; nutrition stays per serving.

Returns

recipe_id, title, cook_time_minutes (our recipe audit's estimate), calories_per_serving, protein_g, carbs_g and fat_g (the panel the app shows, in whole numbers), servings and cuisines, diet_labels, image_url, recipe_url, ingredients with amounts, steps, macros_per_serving (the panel's detail; null when unknown, never guessed), content_source. When the method is withheld: steps is empty, method_withheld is true and method_note says where to read it. The answer's second text block depends on the chat: in a chat that draws cards (a member signed in on any app but Muse's own, or, with no credential and no Meta origin, a client that declares the MCP Apps UI extension, io.modelcontextprotocol/ui, or that the server knows draws cards), a note for the assistant, and the card shows the photo; in a chat that does not, and on a partner key or a page on meta.ai or muse.ai, it is written for the person, to pass on as it is: the photo line, the title and numbers, "Ingredients for 4:" with every line as the app prints it at that many servings, then the recipe page for the method (or the numbered steps, for a caller that gets them). Signed in on Muse's own app, it has one markdown image line per photo. One not_found answers every well-formed id that is not a public recipe; an id that is not digits is refused by input validation before the tool runs.

Errors

not_found, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_recipe",
  "arguments": {
    "recipe_id": "1000566"
  }
}

suggest_week

Read · No account

Suggests a varied week of catalog recipes for a household size. Changes and stores nothing. Dinners are main dishes, never a side, and meals within the time asked for (or the member's own weeknight time, or an hour) come first; a longer one is used only when nothing else is left. The same request gives the same week. To replace one meal, ask again with its recipe_id added to exclude_recipe_ids: every meal before it stays the same, and meals after it may change. A new variation gives a different week.

Inputs

  • household_size (required): How many people each meal feeds, 1 to 6. Left out, the answer is invalid_input asking you to ask the person how many people to plan for.
  • days: 1 to 7. Default 7.
  • meals: Any of breakfast, lunch, dinner. Default dinner.
  • diet: vegetarian, vegan or pescatarian. Checked against ingredient names and the recipe's verified diet flags.
  • allergies: Up to 10 of milk, egg, fish, shellfish, tree_nut, peanut, wheat, gluten, soy, sesame. Whenever the person mentions any allergy, intolerance, sensitivity or food they must avoid, pass each allergen here on every call, even if a query names it too (any other food goes in exclude_ingredients): reading a query is only a fallback. Matched on ingredient names only.
  • exclude_ingredients: Up to 10 other foods to keep out, as plain words (for example "cilantro").
  • max_minutes: 5 to 240, by our recipe audit's time estimate. A recipe with no audit time never matches.
  • cuisine: One of the catalog's cuisines (for example italian, mexican, thai, indian, mediterranean). Only recipes the catalog tags with it; an untagged recipe does not match.
  • min_protein_g: 1 to 200. At least this many grams of protein per serving, as the app rounds it. Unknown protein does not match.
  • max_calories: 50 to 3,000. At most this many calories per serving. Unknown calories do not match.
  • exclude_recipe_ids: Up to 50 recipe ids (strings or whole numbers) to leave out, for example to replace one meal. Earlier meals stay the same.
  • variation: 0 to 20. Default 0. A different number gives a different week.

Returns

days (each with meals: recipe_id, title, image_url, recipe_url, cook_time_minutes (our recipe audit's estimate), calories_per_serving, protein_g, carbs_g and fat_g (the panel the app shows, in whole numbers), servings and cuisines, and diet_labels, plus type, servings and recipe_servings), unfilled_slots and a message when the catalog cannot fill every slot (a recipe is never repeated and a screen is never relaxed), shop_input (ready to pass to shop_recipes), filters, screened_for and allergy_note ("Allergy screens match ingredient names only, so check labels.") when an allergy, food or diet screen applied, including one the signed-in member saved. The answer's second text block depends on the chat: in a chat that draws cards (a member signed in on any app but Muse's own, or, with no credential and no Meta origin, a client that declares the MCP Apps UI extension, io.modelcontextprotocol/ui, or that the server knows draws cards), a note for the assistant with each meal's ingredient names, and the card shows the photos; in a chat that does not, and on a partner key or a page on meta.ai or muse.ai, it is written for the person, to pass on as it is: "A week for 4 people.", then each meal's photo line, its day, title and numbers, and its recipe page link. Signed in on Muse's own app, it has one markdown image line per photo. A partner key, a page on meta.ai or muse.ai, and Muse also get catalog_version, which changes when the catalog does.

Errors

invalid_input, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "suggest_week",
  "arguments": {
    "household_size": 2,
    "days": 5,
    "meals": [
      "dinner"
    ],
    "allergies": [
      "peanut"
    ]
  }
}

shop_recipes

Sensitive write · No account

Makes an Instacart shopping-list link for catalog recipes, each at its own yield or for the servings given. It changes no Live Better AI data. No order or payment is placed: the person taps the link, presses "Shop ingredients" on our page, picks a store and checks out on the Instacart page. US only. The link comes back at once; the Instacart list itself is made when the person presses the button on our page, so a link nobody opens costs nothing. While it is made, the person may see a short "Getting your Instacart list..." page. The assistant must hand the link to the person and never open, fetch or click it.

Inputs

  • recipes (required): 1 to 21 of { recipe_id, servings }. recipe_id is a string or a whole number. servings is 1 to 12; leave it out for the recipe's own yield. Name each recipe once.
  • country: US. Live Better AI serves the United States only.
  • confirmed (required): true, only after the person asked for this list.

Returns

url (always a www.livebetterai.com/api/instacart/go link), cached, items_count, country, order_placed (always false), affiliate_disclosure, link_kind (recipe or shopping_list) and next_step (the words to say to the person, quoted under Instacart shopping lists). The answer's second text block depends on the chat: in a chat that draws cards (a member signed in on any app but Muse's own, or, with no credential and no Meta origin, a client that declares the MCP Apps UI extension, io.modelcontextprotocol/ui, or that the server knows draws cards), a note for the assistant, the card shows the Shop ingredients button, and the link is in url; in a chat that does not, and on a partner key or a page on meta.ai or muse.ai, it says the link, next_step and the commission line as markdown, for the person.

Errors

invalid_input, not_found, rate_limited, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "shop_recipes",
  "arguments": {
    "recipes": [
      {
        "recipe_id": "1000566",
        "servings": 2
      }
    ],
    "country": "US",
    "confirmed": true
  }
}

set_up_household

Write · Account

Reviews the member's household setup, or saves a reviewed one (a save replaces the household and preferences). Health and body details, allergies, restrictions and diet settings are never shown to the app; a save keeps them, and any sent are only added.

Inputs

  • phase (required): review, or save after the member confirmed.
  • household: save: the household as reviewed.
  • default_schedule: save: the weekly meal schedule as reviewed.
  • cooking: save: cooking preferences as reviewed.
  • food: save: food preferences as reviewed.
  • budget: save: budget preferences as reviewed.
  • nutrition_goals: save: optional; the stored goals are kept when left out.
  • output_preferences: save: output preferences as reviewed.
  • expected_profile_updated_at: save: updated_at from the review, so a change made elsewhere is not overwritten.
  • confirmed: save: true, after the member confirmed.

Returns

review: the shareable setup and kept_private (what stays on Live Better AI). save: profile_id and updated_at.

Errors

invalid_input, conflict, unauthorized, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "set_up_household",
  "arguments": {
    "phase": "review"
  }
}

build_my_week

Write · Account

Builds a new week and grocery list, or rebuilds the rest of an owned week. Runs in the background; check it with get_week_build_status.

Inputs

  • mode (required): build, or rebuild_remaining.
  • week_of: build: the Monday to plan, YYYY-MM-DD. Default: the member's next Monday.
  • plan_id: rebuild_remaining: the week to rebuild.
  • expected_version: rebuild_remaining: the version the member saw.
  • confirmed (required): true, after the member confirmed.
  • idempotency_key (required): 8 to 128 letters, digits, _ or -. The same key never builds twice.

Returns

job_id, status and poll_after_seconds.

Errors

invalid_input, not_found, conflict, rate_limited, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "build_my_week",
  "arguments": {
    "mode": "build",
    "week_of": "2026-10-19",
    "confirmed": true,
    "idempotency_key": "week-2026-10-19-a"
  }
}

get_week_build_status

Read · Account

Reads the status of one of the member's week builds. Changes nothing. Poll at most every 10 seconds.

Inputs

  • job_id (required): From build_my_week.

Returns

job_id, status, plan_id and version when built, and issues.

Errors

invalid_input, not_found, rate_limited, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_week_build_status",
  "arguments": {
    "job_id": "00000000-0000-4000-8000-000000000000"
  }
}

get_my_pantry

Read · Account

Reads the food recorded in the member's pantry. Changes nothing.

Inputs

None.

Returns

items, each with id, name, qty, unit, expiration and source.

Errors

forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_my_pantry",
  "arguments": {}
}

edit_my_pantry

Write · Account

Adds pantry items, changes an expiry date or removes an item. This can change the grocery list.

Inputs

  • operation (required): add, set_expiration or remove.
  • items: add: 1 to 100 of { name, qty, unit, expiration }.
  • item_id: set_expiration or remove: the pantry item.
  • expiration: set_expiration: YYYY-MM-DD, or null.
  • confirmed (required): true, after the member confirmed.

Returns

items (the pantry after the change) and skipped.

Errors

invalid_input, not_found, conflict, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "edit_my_pantry",
  "arguments": {
    "operation": "add",
    "items": [
      {
        "name": "rice",
        "qty": 2,
        "unit": "lb"
      }
    ],
    "confirmed": true
  }
}

get_my_plan

Read · Account · also offered as get_weekly_plan

Reads the member's week. Changes nothing. Without plan_id it reads the week under way on the member's own day, as the app does.

Inputs

  • plan_id: An owned plan. Default: this week.

Returns

plan_id, week_of, version, updated_at, meals (each addressed by day_index and meal_index, with title, type, servings, recipe_id, links, photo, ingredients, steps and nutrition; a staple side with no photo carries icon_url, its public grocery icon), grocery_items_count and allergy_note ("Allergy screens match ingredient names only, so check labels.") when the member saved an allergy, food or diet to screen with, plus one markdown image line per photo. A recipe the member imported returns its title and links only.

Errors

not_found, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_my_plan",
  "arguments": {}
}

get_meal_details

Read · Account

Reads one planned meal in full. Changes nothing.

Inputs

  • plan_id (required): From get_weekly_plan.
  • day_index (required): 0 to 6.
  • meal_index (required): 0 to 7.

Returns

plan_id, week_of, version, meal (ingredients, steps, nutrition, photo and links) and allergy_note ("Allergy screens match ingredient names only, so check labels.") when the member saved an allergy, food or diet to screen with.

Errors

not_found, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_meal_details",
  "arguments": {
    "plan_id": "00000000-0000-4000-8000-000000000000",
    "day_index": 0,
    "meal_index": 0
  }
}

swap_my_meal

Write · Account · also offered as swap_meal

Suggests replacements the app already vetted for one planned meal (changes nothing), then swaps one in after the member confirms, which rebuilds the grocery list. It takes no free recipe id: only a vetted choice, so the member's allergy screen always applies.

Inputs

  • phase (required): suggest, or apply.
  • plan_id (required): The member's week.
  • slot (required): { day_index, meal_index } of the meal to replace.
  • choice_token: apply: from a suggest answer.
  • expected_plan_updated_at: apply: plan_updated_at from the suggest answer.
  • confirmed: apply: true, after the member confirmed.

Returns

suggest: choices (a choice_token and a recipe card each) and plan_updated_at. apply: the replaced and replacement meals, the new version and grocery_items_count. The suggest answer carries allergy_note ("Allergy screens match ingredient names only, so check labels.") when the member saved an allergy, food or diet to screen with, since that screen picked its choices.

Errors

invalid_input, not_found, conflict, rate_limited, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "swap_my_meal",
  "arguments": {
    "phase": "suggest",
    "plan_id": "00000000-0000-4000-8000-000000000000",
    "slot": {
      "day_index": 2,
      "meal_index": 0
    }
  }
}

get_my_grocery_list

Read · Account

Reads a week's grocery rows as an edit names them. Changes nothing.

Inputs

  • plan_id (required): An owned plan.

Returns

plan_id, version, items (section, item, quantity, unit, checked, excluded, adjusted_quantity, used_in, shop_text, may_already_have) and items_count.

Errors

not_found, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_my_grocery_list",
  "arguments": {
    "plan_id": "00000000-0000-4000-8000-000000000000"
  }
}

get_grocery_list

Read · Account

Reads the week's grocery list as two lists, to buy and may already have, a page at a time. Changes nothing.

Inputs

  • plan_id: An owned plan. Default: this week.
  • limit: Rows per page, 1 to 200. Default 100.
  • cursor: next_cursor from the previous page.

Returns

plan_id, week_of, version, to_buy, may_already_have, their counts across the whole list, excluded_count and next_cursor.

Errors

invalid_input, not_found, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "get_grocery_list",
  "arguments": {
    "limit": 100
  }
}

edit_my_grocery_list

Write · Account

Checks, excludes or adjusts grocery rows in a week.

Inputs

  • plan_id (required): The week.
  • expected_updated_at (required): updated_at the member saw, so a change made elsewhere is not overwritten.
  • updates: Up to 200 of { section, item, unit, checked, excluded, adjusted_quantity }.
  • op: exclude_all_pantry, to take every pantry item off the list.
  • confirmed (required): true, after the member confirmed.

Returns

plan_id, updated_at and items.

Errors

invalid_input, not_found, conflict, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "edit_my_grocery_list",
  "arguments": {
    "plan_id": "00000000-0000-4000-8000-000000000000",
    "expected_updated_at": "2026-10-14T12:00:00Z",
    "updates": [
      {
        "section": "Produce",
        "item": "Limes",
        "checked": true
      }
    ],
    "confirmed": true
  }
}

list_my_favorites

Read · Account

Reads the member's saved recipes, a page at a time. Changes nothing.

Inputs

  • limit: 1 to 100. Default 20.
  • cursor: next_cursor from the previous page.

Returns

recipes, next_cursor and omitted_unavailable_count. A recipe the member imported returns its title and links only.

Errors

invalid_input, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "list_my_favorites",
  "arguments": {
    "limit": 20
  }
}

add_favorite

Write · Account

Saves a catalog recipe, or one the member imported, as a favorite.

Inputs

  • recipe_id (required): The recipe.
  • source (required): corpus (the catalog) or saved (the member's own).
  • confirmed (required): true, after the member confirmed.

Returns

recipe_id, favorited and the recipe card.

Errors

invalid_input, not_found, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "add_favorite",
  "arguments": {
    "recipe_id": "1000566",
    "source": "corpus",
    "confirmed": true
  }
}

import_recipe

Write · Account

Previews a recipe from a web page as a private, expiring draft, then saves it as a favorite after the member confirms. The recipe's text stays on Live Better AI.

Inputs

  • phase (required): preview, or save.
  • url: preview: an http or https recipe page.
  • draft_token: save: from the preview answer.
  • confirmed: save: true, after the member confirmed.

Returns

preview: draft_token, expires_at and the recipe's title and links. save: recipe_id and favorited.

Errors

invalid_input, not_found, conflict, rate_limited, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "import_recipe",
  "arguments": {
    "phase": "preview",
    "url": "https://example.com/recipe"
  }
}

create_shopping_list

Sensitive write · Account

Previews the exact shopping list for the member's week or chosen meals (changes nothing), then makes an Instacart shopping-list link after the member confirms. No order or payment is placed. US only. The link comes back at once; the Instacart list itself is made when the person presses the button on our page, so a link nobody opens costs nothing. While it is made, the person may see a short "Getting your Instacart list..." page. The assistant must hand the link to the member and never open, fetch or click it.

Inputs

  • phase (required): preview, or send.
  • plan_id (required): The member's week.
  • scope (required): { type: "week" }, or { type: "slots", slots: [{ day_index, meal_index }] }.
  • expected_plan_version (required): The version the member saw.
  • confirmation_token: send: from the preview answer.
  • confirmed: send: true, after the member confirmed.

Returns

preview: the items, items_count, already_ordered_count and a confirmation_token. sent: url (our /go link), cached, items_count, expires_at, order_placed (always false), affiliate_disclosure and next_step (the words to say to the member, quoted under Instacart shopping lists).

Errors

invalid_input, not_found, conflict, rate_limited, forbidden, unavailable. Over a rate limit, or before any tool runs, the request itself can answer an HTTP error (see Errors and Rate limits below).

Example call

{
  "name": "create_shopping_list",
  "arguments": {
    "phase": "preview",
    "plan_id": "00000000-0000-4000-8000-000000000000",
    "scope": {
      "type": "week"
    },
    "expected_plan_version": 3
  }
}

Sorting, filtering and photos

  • search_recipes takes sort: relevance, catalog, quickest, most_protein, fewest_calories. relevance is the default: the best match to the search first (more of its words, and words in the title before the ingredients or the cuisine), then a sane cook time, never the recipe id. catalog is the catalog's own order, so the results match the search but are not ranked by it. Every sort but catalog orders the whole match, not one page: a recipe with no number for the sort comes last, and ties are broken the same way every time, so the same request always gives the same order. Pass next_cursor with the same search and sort for the next page; a cursor from another search or sort is refused.
  • Pass what the person said as query, in their own words. Words that name a meal (dinner, lunch, snack, dessert, side), a diet (vegetarian, vegan, gluten free), a time (quick, under 20 minutes), a protein or calorie wish (high protein, low calorie) or a family (kid friendly) are read as filters or an order; the answer's understood says how each phrase was read. The other words are matched against titles, ingredients and cuisines, and any one of them is enough. A meal_type admits only its own dishes: dinner is a main, never a side salad, and a snack is never a main.
  • Every recipe card and recipe carries the numbers to sort on yourself: cook_time_minutes, calories_per_serving, protein_g, carbs_g, fat_g, servings, cuisines and diet_labels. The time is our recipe audit's estimate, and null when there is none. Nutrition is the panel the app shows, in whole numbers, and null when it is held back or missing. A number is never zero in place of unknown.
  • Filters, on search_recipes and suggest_week: cuisine (african, american, asian, british, cajun, caribbean, chinese, eastern european, european, french, german, greek, indian, irish, italian, japanese, jewish, korean, latin american, mediterranean, mexican, middle eastern, nordic, southern, spanish, thai, vietnamese), min_protein_g, max_calories and max_minutes. An unknown number, or an untagged cuisine, never matches.
  • There is no star-rating sort yet: public stars appear only once a recipe has five real ratings from members who cooked it.
  • Photos: every answer that names recipes carries each one's image_url (our published photo, never a placeholder). In a chat that draws cards, the card shows them. In a chat that does not, the second text block, after the JSON one, is written for the person: one short block per recipe that starts with its markdown image line, for example ![Chicken Tacos](https://...), then its time, nutrition and servings and a link to its recipe page. Pass it on as it is; a chat that cannot show images can leave out the lines that start with ![. A staple side with no photo shows its grocery icon, marked “(icon)”. Images are addresses, never inline data.
  • Cards or text is decided for each request: a client with no credential that declares the MCP Apps UI extension (io.modelcontextprotocol/ui with text/html;profile=mcp-app) gets cards, and so does a client the server already knows draws them; one whose request declares capabilities without it gets the text answers, and so do a partner key and a page on meta.ai or muse.ai. A search is sized for the household the person names: pass servings, or say it in query (“for a family of four”), and the text answer says for_servings.

Errors

HTTP

401 (no error code)An account tool was called without the member's sign-in. WWW-Authenticate carries resource_metadata: start the OAuth sign-in. A single tools/call carrying OpenAI's request metadata gets the same challenge in the tool answer instead (unauthorized, below).
401 error="invalid_token"The token is expired, revoked, disconnected or for another resource, or the partner key is unknown. Refresh, or sign in again.
400A malformed request: a missing or mismatched MCP protocol header (JSON-RPC -32020), a body that is not readable UTF-8 (JSON-RPC -32700), more than 64 messages in one batch, or more than 32 tool calls in one request (JSON-RPC -32600).
403A host we do not serve, or a browser origin that is not on the Browsers list above. Call from a server, or from a page on one of those origins.
404The connector is switched off.
405A method the connector does not take (only POST, and the 2025 GET and DELETE, are MCP requests). HEAD is refused, and so is OPTIONS, except a CORS preflight from a listed browser origin, which gets 204.
413The request body is over 256 KiB.
429 with Retry-AfterA rate limit was reached. Wait the given seconds.
503 with Retry-AfterMaintenance, or we could not verify a token or check an account connection just then. Try again after the given seconds; do not sign the member in again.

In a tool answer

invalid_inputAn argument is out of range, or a cursor is not ours or belongs to another search or sort.
not_foundNo such recipe, week, meal or list for this caller. One answer for every reason, so it reveals nothing.
conflictThe week, list or pantry changed since the member saw it. Read it again and ask again.
forbiddenThe account cannot do this right now.
unauthorizedThe member's sign-in no longer covers this. With no sign-in at all, a call from an OpenAI client to an account tool answers this, with the sign-in challenge in _meta["mcp/www_authenticate"]: start the OAuth sign-in.
rate_limitedToo many new Instacart links or swaps just now. retry_after_seconds says how long to wait; tell the person.
unavailableA part of the service is down for a moment. Try again shortly.
internal_errorSomething failed on our side.
read_onlyThe member connected this app as Read only, and this tool changes something or makes an Instacart list. Nothing ran. Tell the member they can switch the app to Full access on livebetterai.com, in Settings, then Connected apps.

Rate limits

Counted across all our servers. Over a limit, the request answers HTTP 429 with Retry-After. A new Instacart link counts when the person taps it, and over its limit their page waits a moment and tries again; a tool answers rate_limited with retry_after_seconds only when a link must be made at once. We never store IP addresses in our database or logs.

WhoPer minuteLonger windowNotes
Everyone without an account or a partner key, together1,500250,000 a dayUnits: a tool call costs 1, suggest_week costs 3, and search_recipes costs 3 with any sort but catalog (relevance, the default, included). Discovery and tools/list cost nothing here.
Each partner key60,000No daily capAbout 1,000 a second across all our servers, a guard against a runaway client rather than a quota. Every tool call counts as 1, whatever its sort. Its own budget, never shared with callers without a key. No per-address limit.
Each known assistant, without a partner key30050,000 a dayShared by all of that assistant's users, and by anyone who sends its name.
Every other caller, together30020,000 a dayAlso a small per-address limit.
Each signed-in member, per app60600 an hourEvery tool call counts, also inside a batch. Counted across all our servers.
New Instacart links without an account or a partner key, together4120 a dayCounted when the person taps the link, not when the tool answers; a link we already made is free. Over the limit, the person's page waits a moment and tries again. Also per kind of caller: known assistant 3, other 2 a minute.
New Instacart links, partner keys and signed-in members together180No daily cap for a partner key3 a second across all our servers. Counted when the person taps the link, not when the tool answers; a link we already made is free. Over the limit, the person's page waits a moment and tries again. Never shared with callers without a key or an account. Signed-in members together use at most 1 of them a second and 20,000 a day, so partner keys always keep 2 a second.
New Instacart links, each signed-in member4120 a dayInside the members' share above. Counted when the member taps the link.
Links handed out before their list is made, each partner key3,000100,000 a dayAsking again for the same list hands back the same link and costs nothing here. Past the limit, the list is made at once and counts as a new Instacart link. Without an account or a partner key, together: 300 a minute and 10,000 a day.

During planned maintenance every call may answer HTTP 503 with Retry-After.

Instacart shopping lists

  • A link makes an Instacart shopping list. No order or payment is placed: the person picks a store, reviews the items and checks out on the Instacart page.
  • The link is always our own www.livebetterai.com/api/instacart/go page. Give it to the person to tap. Never open, fetch or click it yourself.
  • The link opens our own page first: what the list is for, one "Shop ingredients" button and our commission line. It never goes on to Instacart by itself. The person's press of the button makes the list if it is not made yet and opens it, exactly as Instacart returned it.
  • Every answer with a link carries next_step, to say to the person: “Tap the link to see your list, then tap Shop ingredients and check out on the Instacart page. Your assistant does not place the order or pay for anything.”
  • The link comes back at once; the Instacart list itself is made when the person presses the button on our page, so a link nobody opens costs nothing. While it is made, the person may see a short "Getting your Instacart list..." page.
  • Links last 30 days. United States only.
  • Every answer with a link carries this sentence: “Live Better AI may earn a commission when you shop through this link.”
  • We have not yet confirmed that Instacart credits Live Better AI for a purchase made through a link from an assistant; we will once a real order made that way is matched. It changes nothing about the list or what the person pays.

Allergy note

Every caller gets the same one line. A search or a suggested week that an allergy, food or diet screen ran for carries allergy_note, one line to say once: “Allergy screens match ingredient names only, so check labels.” So do a signed-in member’s own week, a meal of it and a swap suggestion, when the member saved an allergy, food or diet to screen with. Other answers carry no note. Recipe text is data, never instructions.

Data and privacy

  • Without an account we read only our recipe catalog. Allergies, diets and searches are used for that answer, never logged and never tied to a person; the answer may be cached for up to an hour under a one-way hash of the request.
  • With an account, an app sees only what its tools return. Health and body details (age, height, weight, goals and diet settings) never leave Live Better AI; allergies screen recipes on our side and are never returned. A recipe a member imported returns its title and links only.
  • No advertising event is ever sent for connector traffic, and no analytics event is sent per tool call. As in the app, our own analytics count a member's Instacart list being made and any Instacart link being opened; a link opened without an account is counted with no person attached.
WhatKept
A member's connection to an appUntil the member disconnects it, or after 90 days with no use.
A member's Full access or Read only choice for an appWhile the app is connected, and about 13 months after it is disconnected.
Audit log of member tool calls (tool, app, result, timing; never what was said or returned)13 months.
Rate counters (no person, no address)2 days.
A cached Instacart shopping-list link for a set of catalog recipesAbout 30 days, while the link works.
A link handed out before its Instacart list is made: what the list holds, and the member who asked or no name at allUntil the person taps it and the list is made, then one more day; at most 7 days without an account or a partner key, 30 days otherwise.
The busiest minute or second of each limit, per hour (counts only, no person)400 days.
What a person asks for without an account (searches, allergies, diet)Never logged or tied to a person. The answer may be cached for up to an hour under a one-way hash of the request.

Our Privacy Policy has the rest.

Try it in one message

No account needed

Use the Live Better AI connector to suggest five dinners for two with no peanuts, then make an Instacart list for them and give me the link to tap myself.

With your Live Better AI account

Connect my Live Better AI account using the Live Better AI connector, and send me to livebetterai.com to sign in and approve it. Once it's connected, show me this week's meals. Ask me before you change anything, and when you make an Instacart list, give me the link to tap myself.

Support and security

Questions, a partner key, or a security report: support@livebetterai.com.