mealie-mcp
A Model Context Protocol (MCP) server for Mealie, the self-hosted recipe manager, meal planner and shopping-list app.
It exposes every endpoint of the Mealie REST API as an MCP tool, so an LLM (Claude, etc.) can read and manage your recipes, meal plans, shopping lists, cookbooks, households, users and more.
- 🧩 Broad API coverage, sane baseline — one tool per Mealie endpoint,
trimmed to a safe baseline (~211 of ~259) so clients aren't overwhelmed and risky endpoints aren't reachable.
- 🔄 Auto-adapts to your Mealie version — on startup it fetches the OpenAPI
schema from your instance, so the tools always match exactly what your server supports. A bundled snapshot is used as a fallback if the fetch fails.
- 🚀 Zero install — runs straight from
npx, ideal for MCPHub, Claude
Desktop, Cursor, and any other MCP client.
- 🔒 Safe by default — admin/server-ops endpoints are never exposed, plus
read-only mode and per-category include/exclude filtering to narrow further.
- 🔑 Flexible auth — a static API token, or an OAuth2 client-credentials flow
that fetches and refreshes access tokens for you.
---
Quick start
You need two things:
- The base URL of your Mealie instance, e.g.
https://mealie.example.com - A Mealie API token (see Getting an API token)
Run it with npx:
MEALIE_BASE_URL="https://mealie.example.com" \
MEALIE_API_TOKEN="your-long-lived-token" \
npx -y mealie-mcp
The server speaks MCP over stdio, so you normally won't run it by hand — your MCP client launches it for you using the config below.
---
Client configuration
MCPHub / Claude Desktop / Cursor (generic MCP config)
Add an entry to your client's MCP servers config (for Claude Desktop this is claude_desktop_config.json; MCPHub uses an equivalent mcpServers block):
{
"mcpServers": {
"mealie": {
"command": "npx",
"args": ["-y", "mealie-mcp"],
"env": {
"MEALIE_BASE_URL": "https://mealie.example.com",
"MEALIE_API_TOKEN": "your-long-lived-token"
}
}
}
}
That's the only required configuration. Everything else is optional tuning.
---
Getting an API token
In Mealie:
- Click your user avatar → Manage Your Profile.
- Open the API Tokens section (
/user/profile/api-tokens). - Create a token, give it a name, and copy it.
The token inherits the permissions of the user that created it, so create it under a user/household with the access you want the LLM to have. To give the model read-only-ish safety, also see MEALIE_READ_ONLY.
---
Configuration
All configuration is via environment variables.
| Variable | Required | Default | Description | | --- | --- | --- | --- | | MEALIE_BASE_URL | ✅ | — | Base URL of your Mealie instance, e.g. https://mealie.example.com. | | MEALIE_API_TOKEN | – | — | Long-lived Mealie API token (sent as Authorization: Bearer). Most endpoints need it. MEALIE_TOKEN is accepted as an alias. Ignored when OAuth is configured. | | MEALIE_OAUTH_TOKEN_URL | – | — | IdP token endpoint. Setting this (plus client id/secret) enables the OAuth2 client-credentials flow, which takes precedence over MEALIE_API_TOKEN. See Authenticating with OAuth. | | MEALIE_OAUTH_CLIENT_ID | – | — | OAuth client id (required for the OAuth flow). | | MEALIE_OAUTH_CLIENT_SECRET | – | — | OAuth client secret (required for the OAuth flow). | | MEALIE_OAUTH_SCOPE | – | — | Optional space-delimited OAuth scopes. | | MEALIE_OAUTH_AUDIENCE | – | — | Optional OAuth audience (some IdPs, e.g. Auth0, need it to mint a Mealie-targeted token). | | MEALIE_READ_ONLY | – | false | When true, only expose GET endpoints. Great for a safe, read-only assistant. | | MEALIE_TOOLS | – | — | Comma-separated allow-list of tool names or category slugs to expose (e.g. recipe,households_shopping_lists). Empty = the full safe baseline. | | MEALIE_EXCLUDE_TOOLS | – | — | Comma-separated deny-list of tool names or category slugs to hide further (e.g. groups_seeders,groups_migrations). Applied on top of the always-on baseline trim. | | MEALIE_USE_BUNDLED_SPEC | – | false | Skip the live OpenAPI fetch and use the snapshot bundled with the package. | | MEALIE_OPENAPI_URL | – | ${MEALIE_BASE_URL}/openapi.json | Override where the OpenAPI schema is fetched from. | | MEALIE_TOOL_NAME_MAX | – | 50 | Max length of generated tool names (clamped to 16–64). Lower it if your MCP client prefixes tool names (e.g. mcp__<server>__<tool>) and the combined name exceeds the 64-char API limit. | | MEALIE_TIMEOUT | – | 60000 | Per-request timeout in milliseconds. | | MEALIE_RETRIES | – | 2 | Extra attempts for idempotent (GET) requests that hit a network error or a retryable status (429/5xx), with exponential backoff (clamped to 0–5). Non-GET methods are never retried automatically. Set 0 to disable. | | MEALIE_DEBUG | – | false | When true, log each outgoing request (method, path, response status) to stderr. Useful for troubleshooting from your MCP client. | | MEALIE_ACCEPT_LANGUAGE | – | — | Optional Accept-Language header forwarded to Mealie (affects e.g. ingredient parsing locale). | | MEALIE_ALLOWED_UPLOAD_DIRS | – | — | Comma-separated directories that file uploads may read from. Unset (the default) means uploads are disabled. See Restricting file uploads. |
Note: Only path and query parameters are exposed as tool inputs. The few Mealie endpoints that read a custom request header or cookie are not driven through those parameters —
Accept-Languageis forwarded viaMEALIE_ACCEPT_LANGUAGE, and authentication is handled globally.
Restricting file uploads
A handful of Mealie endpoints take a file — recipe images, recipe assets, ZIP imports, backups. For those, the tool argument is a path on the machine running this server, and the server reads that path and sends the bytes to Mealie.
That means anything the server process can read, a tool call can upload. Since the tool call is chosen by a model, and models read recipe pages and other untrusted text, a prompt-injection payload can in principle ask for a path you never intended — ~/.ssh/id_rsa rather than a photo of dinner.
MEALIE_ALLOWED_UPLOAD_DIRS bounds that. Set it to a comma-separated list of directories, and any upload resolving outside all of them is refused:
MEALIE_ALLOWED_UPLOAD_DIRS="/home/me/Pictures/recipes,/srv/mealie/imports"
Paths are compared after symlinks are resolved on both sides, so a symlink inside an allowed directory cannot point out of it, a symlinked allowed directory still works, and /srv/mealie/imports-scratch is not treated as being inside /srv/mealie/imports. If the variable is set but none of its directories exist, every upload is refused rather than silently falling back to unrestricted.
Leaving it unset disables all uploads for security. If you use upload tools, setting this variable is required. If you never upload files, consider dropping those tools entirely with MEALIE_EXCLUDE_TOOLS, or run the server read-only with MEALIE_READ_ONLY=true.
Authenticating with OAuth (client credentials)
By default the server authenticates with a static MEALIE_API_TOKEN. As an alternative — useful for headless/machine-to-machine setups — it can obtain an access token from your identity provider using the OAuth2 client-credentials grant, then send it as the Authorization: Bearer credential and refresh it automatically (proactively before expiry, and reactively on a 401).
MEALIE_OAUTH_TOKEN_URL="https://idp.example.com/oauth/token"
MEALIE_OAUTH_CLIENT_ID="your-client-id"
MEALIE_OAUTH_CLIENT_SECRET="your-client-secret"
# Optional, IdP-dependent:
MEALIE_OAUTH_SCOPE="mealie"
MEALIE_OAUTH_AUDIENCE="https://mealie.example.com"
When these are set, OAuth takes precedence and MEALIE_API_TOKEN is ignored.
Precondition: your Mealie must be a version that validates IdP-issued access tokens as bearer tokens (via the provider's JWKS). The client sends the credentials in the request body (
client_secret_post). If your IdP requires HTTP Basic client auth instead, open an issue.
The safe baseline (≈211 of 259 endpoints)
Mealie exposes ~259 endpoints, but the server ships a safe baseline of ~211 so clients aren't overwhelmed and risky endpoints aren't reachable at all. Two groups are permanently excluded — they can't be re-enabled by configuration:
- Admin / server-ops endpoints (backups, maintenance, multi-tenant
user/group/household management, debug, email config, AI providers). These are powerful, rarely what a recipe assistant needs, and a security footgun, so the server never exposes them. Manage your instance through Mealie's own UI/API.
- Endpoints with little value to an LLM: password-reset and registration flows,
the docker healthcheck route, the SSE stream duplicates of plain-JSON recipe-import endpoints, and zip/file download routes that return opaque bytes. (The Users: Authentication category is kept, since the server can authenticate via OAuth.)
You can narrow further from this baseline — but not widen past it — with MEALIE_TOOLS / MEALIE_EXCLUDE_TOOLS using category slugs (or exact tool names). Examples:
# Only recipes, meal plans and shopping lists:
MEALIE_TOOLS="recipe,households_mealplans,households_shopping"
# Baseline, but also drop group seeders + migrations:
MEALIE_EXCLUDE_TOOLS="groups_seeders,groups_migrations"
# Read-only recipe browsing assistant:
MEALIE_READ_ONLY=true
MEALIE_TOOLS="recipe,explore"
---
How tools are named
Each tool name is derived from the Mealie OpenAPI tag (category) and operation. Uniquely-named operations use their bare operation name (e.g. suggest_recipes); operations whose name is reused across resources (the CRUD verbs get_all, get_one, create_one, update_one, delete_one, …) are prefixed with their category to stay unique and to keep them grouped. All names are kept well under the 64-character tool-name limit. For example:
| Tool | Method & path | | --- | --- | | recipe_crud_get_all | GET /api/recipes | | recipe_crud_get_one | GET /api/recipes/{slug} | | recipe_crud_create_one | POST /api/recipes | | households_shopping_lists_get_all | GET /api/households/shopping/lists | | households_mealplans_create_one | POST /api/households/mealplans | | app_about_get_app_info | GET /api/app/about |
Each tool's input schema declares its path parameters, query parameters and (where relevant) a body object — all generated directly from Mealie's OpenAPI schema, so the model gets accurate, fully-typed arguments.
File uploads
Endpoints that upload files (recipe images, ZIP imports, backups, assets, …) take their file fields as absolute paths to local files, which the server reads and sends as multipart form data. The tool description tells the model which fields are file paths.
Troubleshooting: "name: String should have at most 64 characters"
Some MCP clients/hubs (e.g. MCPHub, remote connectors) prefix every tool name with the server name — mcp__<server>__<tool> — and the combined string must stay within the API's 64-character limit. If you hit this error:
- Keep the server's name/alias short (
mealieis ideal). - Lower the tool-name cap, e.g.
MEALIE_TOOL_NAME_MAX=30, until it fits. - The error is global — Claude rejects the whole tool list if any tool
(from any server) is too long, so the culprit may be a different server.
Categories
<details> <summary>All 57 categories</summary>
| Category | Tools | | --- | --- | | admin_about | 3 | | admin_ai_providers | 4 | | admin_backups | 6 | | admin_debug | 1 | | admin_email | 2 | | admin_maintenance | 5 | | admin_manage_groups | 5 | | admin_manage_households | 5 | | admin_manage_users | 7 | | app_about | 3 | | explore_categories | 2 | | explore_cookbooks | 2 | | explore_foods | 2 | | explore_households | 2 | | explore_recipes | 3 | | explore_tags | 2 | | explore_tools | 2 | | groups_ai_provider_settings | 2 | | groups_ai_providers | 4 | | groups_households | 2 | | groups_migrations | 1 | | groups_multi_purpose_labels | 5 | | groups_reports | 3 | | groups_seeders | 3 | | groups_self_service | 6 | | households_cookbooks | 6 | | households_event_notifications | 6 | | households_invitations | 3 | | households_mealplan_rules | 5 | | households_mealplans | 7 | | households_recipe_actions | 6 | | households_self_service | 7 | | households_shopping_list_items | 8 | | households_shopping_lists | 9 | | households_webhooks | 7 | | organizer_categories | 7 | | organizer_tags | 7 | | organizer_tools | 6 | | recipe_bulk_actions | 8 | | recipe_comments | 6 | | recipe_crud | 23 | | recipe_exports | 2 | | recipe_images_and_assets | 5 | | recipe_ingredient_parser | 2 | | recipe_shared | 2 | | recipe_timeline | 6 | | recipes_foods | 6 | | recipes_units | 6 | | shared_recipes | 4 | | users_authentication | 5 | | users_crud | 6 | | users_images | 1 | | users_passwords | 2 | | users_ratings | 5 | | users_registration | 1 | | users_tokens | 2 | | utils | 1 |
</details>
---
Development
git clone https://github.com/2fst4u/mealie-mcp.git
cd mealie-mcp
npm install
npm run build # compile TypeScript to dist/
npm test # run the test suite (node:test)
npm run typecheck # type-check without emitting
# Run from source against a Mealie instance:
MEALIE_BASE_URL="https://demo.mealie.io" npm run dev
Updating the bundled OpenAPI snapshot
The server prefers the live schema from your own instance, but the bundled snapshot (used as a fallback) can be refreshed from any Mealie instance:
npm run refresh-spec -- https://demo.mealie.io
Project layout
| Path | Purpose | | --- | --- | | src/index.ts | Entry point: load config + spec, start stdio server. | | src/config.ts | Environment-variable configuration. | | src/auth.ts | Resolve the Authorization header (static token or OAuth client credentials). | | src/openapi-loader.ts | Fetch live OpenAPI schema with bundled fallback. | | src/tools.ts | Generate one MCP tool per OpenAPI operation; built-in trimming. | | src/schema.ts | Build self-contained JSON Schemas ($ref → $defs). | | src/http-client.ts | Execute requests (JSON / urlencoded / multipart / binary). | | src/server.ts | MCP server wiring (tools/list, tools/call). | | openapi.snapshot.json | Bundled fallback OpenAPI schema. |
---
Releases & publishing
This repo ships two GitHub Actions workflows:
- CI (
.github/workflows/ci.yml) — runs type-check, build and tests on every
pull request across Node 18/20/22.
- Release (
.github/workflows/release.yml) — on every push/merge tomainthat touches source files,
builds and tests, auto-bumps the patch version, creates a GitHub Release v<version> (with auto-generated notes), and publishes to npm.
Publishing to npm (Trusted Publishing / OIDC)
The release workflow publishes via npm Trusted Publishing, so there is no NPM_TOKEN secret to store — GitHub Actions authenticates to npm with a short-lived OIDC token, and npm generates build provenance automatically.
npm requires a package to exist before you can attach a Trusted Publisher, so there is a one-time bootstrap for the first ever publish:
- Publish the first version manually from your machine (this also claims the
package name): ``bash npm install # ensure deps npm login # your normal npm account + 2FA npm publish --access public ``
- Configure the Trusted Publisher on npmjs.com: open the package page →
Settings → Trusted Publisher → GitHub Actions, and enter:
- Organization or user:
2fst4u - Repository:
mealie-mcp - Workflow filename:
release.yml - Environment: (leave blank)
- From then on, **every merge to
mainthat touches source files auto-bumps the patch version and publishes
automatically** over OIDC — no tokens, no manual steps.
Requirements (handled by the workflow):
id-token: writepermission, Node ≥ 22.14, and npm ≥ 11.5.1. The publish step safely skips if the version is already on npm, so re-runs and the bootstrap version won't cause failures.
---
License
This project is an independent client for Mealie and is not affiliated with the Mealie project.











