Connect Claude to weReci

weReci is a personal recipe collection — the recipes you save from photos, links, PDFs and videos (plus letters and emails you keep alongside them), the shelves an agent curates for you, and the web of connections between ingredients, techniques, and entries that grows as you cook. This connector lets Claude work with that collection directly, so you can ask it things like “what should I make with what's already in my pantry shelf” or “how are my grandmother's bread recipes related to the sourdough one I just saved” without leaving your conversation with Claude.

It runs over the Model Context Protocol's Streamable HTTP transport, with standard OAuth 2.1 (dynamic client registration, PKCE) — the same kind of connection Claude uses for any other remote MCP server. You sign in with your existing weReci account; there's nothing new to set up on our side.

The one-click way — Claude's connector directory

weReci is listed in Claude's connector directory, so the fastest path is one button: open the listing and choose Connect. Claude walks you through the same weReci sign-in and consent screen described below.

Add weReci to Claude

Claude.ai (web) and Claude Desktop — manually

1. 1Open Settings → Connectors and choose Add custom connector.
2. 2Paste the server URL: https://wereci.xyz/api/mcp
3. 3Claude opens weReci's sign-in page. Sign in with your existing weReci account (or create one) and approve the permissions on the consent screen.
4. 4You're connected — ask Claude anything about your collection.

Claude Code

Add the server from the command line:

claude mcp add --transport http wereci https://wereci.xyz/api/mcp

Claude Code will prompt you to authorize in the browser the first time you use a weReci tool — same sign-in and consent screen as above.

weReci's consent screen offers three permissions, and you only ever see what your account already has access to.

• recipes:read (granted by default) — read-only browsing and search across your collection: recipes, curated shelves, the connections graph, and what weReci remembers about you. Nothing in your collection is created, edited, or deleted.
• email:send (optional) — lets Claude email a recipe, or a grocery list you approved, to you. There is no recipient field anywhere in these tools: they can only ever send to the verified email address on your own signed-in account, so a misbehaving or prompt-injected agent has no way to mail anyone else. It's also rate-limited.
• grocery:cart (optional) — lets Claude send a grocery list you've approved to your own linked store: a Kroger-family cart (Fred Meyer, Mariano's, Ralphs…) or Pete's Fresh coupon clipping. Linking a store is a separate one-time sign-in at the store, and items only ever land in your own cart or on your own loyalty card — you review everything at checkout, on the store's site or app.
• list:sync (optional) — lets the connection see the shopping list you share with your cookbook partner and tick, add, or remove lines on it, the way Home Assistant's to-do list does. Only a list you have chosen to share is ever visible; a list kept on your phone stays on your phone.
• cook:assist (optional) — lets a kitchen display that Home Assistant runs with no phone scale a recipe or suggest an ingredient swap while you cook, the way Cook Mode does. It is not a Claude tool: the display calls weReci directly. Each one counts against your AI allowance like the same tap in the app, and nothing is ever saved to your collection.

Two actions need no extra permission because they can only touch things that are already yours and already undoable: putting items on your own in-app shopping list (add_to_shopping_list — entries are labeled with which assistant added them and expire on their own after 14 days), and reshuffling your curated shelves (reshuffle_conceits — it rearranges shelf themes and nothing else; no recipe, letter, or photo is ever modified or lost). On clients that support it, weReci asks you directly before a reshuffle runs — confirm, and optionally give it a creative direction. Elsewhere, its instructions tell Claude to ask you in the chat first.

A connection keeps the permissions it was granted. When a tool needs one it doesn't have, newer MCP clients ask you to approve just that permission on the spot, and the approval replaces the old connection rather than adding a second one. On other clients, disconnect and reconnect to tick the new box.

In plain language, here's what Claude can call once you've connected. The first group only reads; the second group acts, and each action is caged to your own account.

Reading your collection

• ✦search_recipes — search your saved recipes by keyword, ingredient, or a described idea (“something cozy for a rainy night” works).
• ✦get_recipe — fetch one entry's full detail — ingredients, instructions, notes, photo — by id. Letters and emails in your collection come back in the same shape, marked by kind. Renders as an interactive card where Claude supports it (see below).
• ✦list_conceits — list your current curated shelves (themed groups of entries) with their titles, descriptions, covers and members. Renders as a tap-through shelf browser where supported.
• ✦get_graph_node — everything directly connected to one node in the graph, grouped by relationship type — every recipe using an ingredient, every entry from one author.
• ✦traverse_graph — walk the connections graph outward from a starting point.
• ✦find_graph_connection — find how two things in your collection are related — e.g. the path between two recipes or ingredients.
• ✦get_background_job — check on a job started with start_background_job: running, done (with its result), or failed (with the reason). A run that died is reported as failed rather than left spinning; finished jobs are kept for 24 hours.

Acting on your behalf

• ✦reshuffle_conceits — regenerate your curated shelves with a fresh theme, optionally steered by a creative direction you give it. Asks you first; touches only shelf arrangement, never your data. Runs in the background for a minute or so.
• ✦start_background_job — start a long-running job and get a job id to poll. Today the only kind is a shelf reshuffle — the same thing as above, for agents that want a straight answer about whether their run finished. Claude confirms with you before starting one.
• ✦send_recipe_email — email one recipe to your own verified address, with an optional note. Never to anyone else — see Permissions above.
• ✦build_shopping_list — turn one to eight recipes into a single consolidated grocery list — proposed to you first, and only sent after you approve, in a later turn: to your own linked Kroger-family cart, clipped as this week's matching Pete's Fresh coupons (only items currently on special can match, so unmatched items are normal), as one link that fills your own Walmart cart when you open it (no account link needed), or emailed to you to forward to whoever shops. weReci stores no grocery lists — the list is an artifact of the conversation.
• ✦add_to_shopping_list — put recipes you ask for — or one-off items like “paper towels” — on your own shopping list in the app. Only when you ask, only your own list (or your shared list if you have one), labeled with which assistant added them; entries expire on their own after 14 days if you never open them.

Alongside the tools, weReci exposes MCP resources under the wereci:// scheme. They are read-only projections of the same data — Claude can pull one in without a tool call.

• ✦wereci://schema — the data model — graph vocabulary, recipe fields, and which tool or resource answers which kind of question.
• ✦wereci://collection/overview — a facet map of your collection — recipe count, cuisines, dietary classes, source types.
• ✦wereci://graph/overview — a summary of your connections graph — how many entries, ingredients, authors, sources, and how densely they interlink.
• ✦wereci://conceits — your current curated shelves with member ids.
• ✦wereci://recipe/{id} — one entry's full payload by id.
• ✦wereci://graph/node/{id} — one graph node and everything connected to it.
• ✦wereci://cuisine — Wikipedia's cuisine hierarchy — continents, regions, national and regional cuisines — the same tree weReci classifies your recipes against. No personal data.
• ✦wereci://cuisine/{path} — one cuisine in that tree with its Wikipedia summary, where it sits, and its sub-cuisines — e.g. wereci://cuisine/asia/east_asian/korean.
• ✦wereci://memory — “what this cookbook knows about you” — live collection statistics plus the currently-valid facts weReci has learned from your own conversations and cooking: preferences, dietary constraints, occasions, patterns. Each fact records when it became true and where it was learned.
• ✦wereci://memory/{category} — facts in one category — preference, dietary, occasion, pattern, household, skill, culture, correction — or history for facts that were once true and have since been superseded, each pointing at what replaced it.
• ✦wereci://stats — deterministic numbers only — entry, cuisine, dietary and ingredient counts, shelf summary, joint-cookbook share, 30-day activity. No personal facts.
• ✦wereci://list — what the server knows about your shopping list: your live shared list if you have one, or assistant-added items waiting for your app. The personal list on your own device is never readable from here.

On clients that support MCP Apps — claude.ai and Claude Desktop today — two of the tools render as small interactive views instead of raw text: get_recipe shows a recipe card, and list_conceitsshows your shelves as a browser you can tap through, opening any member's card in place.

The views run in a sandboxed frame and only ever display the data the tool call already returned — they hold no credentials of their own. Tapping through to a recipe opens its page on wereci.xyz in a new tab (Claude may ask you to confirm the link first). Clients without MCP Apps support get the same information as plain text.

The connector sees exactly what your account sees in the app — tenant-scoped to you, including a joint cookbook if you share one with someone. It cannot reach another account's collection under any circumstance.

Nothing in the connector can create, edit, or delete a recipe, letter, or photo. The only things it can write are your own shopping list, the arrangement of your own shelves, an email to your own inbox, and items in your own store cart — each only after you ask.

You can revoke access at any time from weReci settings → Connected apps. Revoking immediately invalidates the connection — nothing further is readable until you reconnect.

For the full picture of what weReci collects and how it's used, see the privacy policy and the terms of service.