For people who already talk to Claude or ChatGPT

Your assistant is great at dinner ideas. It just can't remember one.

Close the tab and the plan is gone. Paste this one URL into Claude, ChatGPT, or Cursor and the same assistant gets 19 real tools — it writes the week into a vault that's still there tomorrow, builds the grocery list, and checks things off while you shop.

https://www.mealplanvault.com/mcp

6 read tools, 13 write tools, one sign-in. Free on every plan, including the free one.

Create a vault, free How to connect it

Verified withClaudeClaude CodeChatGPTCursorGooseZed

Three ways to plan a week

A chat can describe your week. It can't keep it.

A spreadsheet remembers but can't think. A chat thread thinks but can't remember. Connect the vault and your agent gets both: somewhere durable to put the answer, and the tools to put it there.

Dimension	A spreadsheetThe one you keep meaning to update	An ongoing chatClaude or ChatGPT, no tools	The vault over MCPSame agent, real tools
Where the week lives	A grid you keep by hand	Somewhere up the scroll	A real database your agent reads and writes
What the assistant knows	Nothing, unless you paste it in	Whatever is still in the context window	It checks the live plan before it answers
Making one change	Find the cell, retype it	Re-describe the week so it has context again	“Move Thursday to Friday.” One tool call
Recipes	Links you hope still resolve	Re-pasted every session	Imported once, structured, searchable forever
The grocery list	Typed from scratch each week	A list in a message you copy somewhere else	Built from the plan, sorted by aisle
At the store	Squinting at a spreadsheet on your phone	Scrolling for the message with the list	Ask what’s left, check items off as you go
The rest of the household	One person’s file, emailed around	One person’s private conversation	Shared plan, real roles, everyone sees it
Next week	Duplicate the tab, edit it down	Start over and re-explain your family	Already there. So is last month, if you ask
Checking its work	You typed it, so you know	It says it did. Nothing actually changed	Every call logged, named, timestamped

A spreadsheet

The one you keep meaning to update

Where the week lives: A grid you keep by hand
What the assistant knows: Nothing, unless you paste it in
Making one change: Find the cell, retype it
Recipes: Links you hope still resolve
The grocery list: Typed from scratch each week
At the store: Squinting at a spreadsheet on your phone
The rest of the household: One person’s file, emailed around
Next week: Duplicate the tab, edit it down
Checking its work: You typed it, so you know

An ongoing chat

Claude or ChatGPT, no tools

Where the week lives: Somewhere up the scroll
What the assistant knows: Whatever is still in the context window
Making one change: Re-describe the week so it has context again
Recipes: Re-pasted every session
The grocery list: A list in a message you copy somewhere else
At the store: Scrolling for the message with the list
The rest of the household: One person’s private conversation
Next week: Start over and re-explain your family
Checking its work: It says it did. Nothing actually changed

The vault over MCP

Same agent, real tools

Where the week lives: A real database your agent reads and writes
What the assistant knows: It checks the live plan before it answers
Making one change: “Move Thursday to Friday.” One tool call
Recipes: Imported once, structured, searchable forever
The grocery list: Built from the plan, sorted by aisle
At the store: Ask what’s left, check items off as you go
The rest of the household: Shared plan, real roles, everyone sees it
Next week: Already there. So is last month, if you ask
Checking its work: Every call logged, named, timestamped

What it looks like in practice

Say the thing. It does the typing.

These aren't prompts for a better answer. They're instructions that change a database — and leave something behind when the conversation ends.

“Plan next week's dinners around what's already in the vault, and keep Tuesday meatless.”

list_recipesget_meal_plan_weekset_meal_plan_entry

Seven slots filled in the shared plan. Your household opens the app and sees the same week you just talked through.

“Turn next week's plan into a grocery list sorted by aisle.”

get_meal_plan_weekget_recipecreate_grocery_list

A real list, built from the ingredients of the recipes actually on the plan — not a guess at what a week of food looks like.

“Import this recipe I just found and put it on Thursday.”

import_recipe_from_urlset_meal_plan_entry

The recipe is parsed into ingredients and steps and saved for good. Next year it's still searchable, link rot or not.

“We doubled the harissa in last night's traybake — save that on the recipe.”

list_recipesupdate_recipe_notes

The change is on the recipe itself, so whoever cooks it next reads it. Notes from a chat window are gone by Tuesday.

“I'm at the store — read me what's left and check things off as I go.”

get_grocery_listcheck_grocery_list_item

Items check off as you walk. Whoever else is shopping sees them disappear in real time.

“Swap Wednesday for something faster that uses the chicken we already have.”

get_meal_plan_weeklist_pantry_itemslist_recipesset_meal_plan_entry

It reads the current week and your pantry before deciding, so it swaps the meal you actually have scheduled for one you can actually cook tonight.

“What's about to expire, and what can I cook with it this week?”

list_pantry_itemslist_recipesset_meal_plan_entry

The food you already paid for ends up on the plan instead of in the bin — because the pantry is a record it can read, not something you have to recite.

“What did we eat most often last month?”

get_meal_plan_weekget_recipe

It answers from your own history, because the history exists. A fresh chat has nothing to look at.

The whole surface

19 tools. No screen-scraping.

These are the same tools the app itself runs on — 6 read, 13 write. Your agent picks which ones to call. You decide which ones it's allowed to.

Recipes

Search the vault, read a full recipe, write new ones, import from a link, note what to change next time.

readlist_recipes
readget_recipe
writecreate_recipe
writeimport_recipe_from_url
writeupdate_recipe_notes

Meal plans

Read a week, schedule a meal into any slot, clear one out.

readget_meal_plan_week
writeset_meal_plan_entry
writeremove_meal_plan_entry

Grocery lists

Build a list from the plan, add and remove items, check things off mid-aisle.

readlist_grocery_lists
readget_grocery_list
writecreate_grocery_list
writeadd_grocery_list_item
writecheck_grocery_list_item
writeremove_grocery_list_item
writedelete_grocery_list

Pantry

See what's already on hand, add what you bought, draw it down as you cook.

readlist_pantry_items
writeadd_pantry_item
writeupdate_pantry_item
writeremove_pantry_item

Delegation only works if you can audit it.

Giving something write access to your week is a real decision. Here's everything you get to hold it to.

One-click OAuth

OAuth 2.1 with PKCE. There are no API keys to paste, and your password never touches the agent.

Your permissions, inherited

An agent gets the same household role you have. If you are a viewer, its agent can read the plan and nothing else.

Approve tool by tool

Clients like Claude list every tool by name and let you set each one to ask first, always allow, or off.

Reads and writes are separate

A read-only connection cannot reach a single write tool. The scope is enforced on the server, not in the client UI.

Every call logged

Reads and writes both land in an activity log in your dashboard, with a timestamp and the name of the agent that made the call.

Revoke in one tap

Disconnect an agent and its tokens stop working immediately. The record of what it did stays.

Create a vault, free Read the connection guide

The questions worth asking first.

Want the setup steps instead? Read the connection guide.

Why not just keep asking ChatGPT or Claude to plan my week?

You can, and it will give you a good week. The problem is what happens next. A chat answer isn't state: nothing was written down, so the plan lives in a message you have to scroll back to, and the model only knows what's still in the context window. Ask for a change on Thursday and you're re-explaining your household, your allergies, and what you already bought. Connecting The Meal Plan Vault over MCP keeps the same conversation but gives it somewhere to put the answer — the plan, the recipes, and the grocery list land in a real database that's still there next week, shared with everyone in your household, and readable by the agent the next time you ask.

Why is this better than the spreadsheet I already have?

A spreadsheet is durable, which is why people keep using them, but it can't do anything on its own. You maintain every cell by hand, the grocery list is retyped each week from the recipes you're holding in your head, and your assistant can't see any of it unless you paste it in. The vault keeps the durability and adds the part a sheet can't do: the recipes are structured, so a grocery list can be generated from what's actually on the plan; the week is shared, so nobody is emailing a file around; and an agent can read and change it in a sentence. Meal planning is the rare chore that's both repetitive and full of judgment calls — a sheet handles the repetition and none of the judgment.

Which AI apps can connect?

Any client that speaks MCP (Model Context Protocol). We've verified Claude — desktop, web, and mobile connectors — plus Claude Code, ChatGPT's custom connectors, Cursor, Goose, and Zed's agent panel. Setup is the same everywhere: paste https://www.mealplanvault.com/mcp into the client, sign in, approve the connection. There are no API keys to generate or paste.

What does the agent actually get access to?

Fourteen tools — five read and nine write — which are the same tools the app itself runs on. It can search and read recipes, create one, import one from a URL, read any week of the plan, schedule or clear a slot, and create, edit, check off, or delete grocery lists. It cannot see households you're not a member of, and it inherits your household role: if you're a viewer, your agent can read the plan and nothing else.

Do I have to trust it with write access?

No — read and write are separate scopes, enforced on the server rather than in the client's UI, so a read-only connection physically cannot reach a write tool. If you do grant writes, clients like Claude list every tool by name and let you set each one to ask first, always allow, or off. Every call, read or write, lands in an activity log in your dashboard with a timestamp and the name of the agent that made it. Revoking a connection kills its tokens immediately, and the log of what it did stays.

Is the MCP server an upgrade or an add-on?

Neither. It's included on every plan, including the free one, at no extra cost. The free plan is a real plan you can stay on — it isn't a timed trial — so you can connect an agent, plan a few weeks, and decide later whether you want the paid features.

Do I have to use an agent for any of this to work?

No. The Meal Plan Vault is a full meal planning app on its own — a drag-and-drop weekly calendar, recipe import, household voting, shared grocery lists. The MCP server is a second front door, not the only one. Most households end up using both: the agent to draft the week in a sentence, the app to nudge things around afterward.