🍽️ plan-to-eat-mcp

Stop typing recipes. Start telling them. Hand your favorite LLM the keys to your Plan to Eat recipe book, planner, and shopping list — over the Model Context Protocol.

A drop-in MCP server, a CLI, and a standalone Node client that give Claude, ChatGPT, any MCP-aware assistant, or your own shell full read/write access to your Plan to Eat account.

You:  "Save the Bon Appétit miso pasta from this URL and tag it weeknight."
LLM:  ✓ created recipe #48,872,094 — 7 ingredients, 25 min, tagged weeknight.

You:  "Plan dinner around it Tuesday and add the missing pantry items to my
       shopping list."
LLM:  ✓ event created 2026-05-05 / dinner / 4 servings.
      ✓ shopping list now reflects miso, mirin, and dashi.

…or, when you'd rather not talk to anything:

$ plan-to-eat get-planner-week 2026-05-04
start_date  2026-05-04
end_date    2026-05-10

ID        DATE        SECTION    KIND        RECIPE_ID  RECIPE_TITLE        DESCRIPTION            SERVINGS
────────  ──────────  ─────────  ──────────  ─────────  ──────────────────  ─────────────────────  ────────
67698872  2026-05-05  breakfast  recipe      17314561   Breakfast Sandwich                         8
67483177  2026-05-05  dinner     ingredient                                 Asparagus (in season)  0
67710784  2026-05-06  dinner     note                                       Cucumber salad         0

3 rows

That's it. No recipe re-typing. No copy-pasting URLs into a phone app. Just talk.

❤️ Why Plan to Eat?

If you're not already using it, Plan to Eat is genuinely the best meal-planning app I've ever used:

• A clip-anywhere recipe importer that actually works on real-world food blogs.
• Drag-and-drop weekly planner that automatically rolls ingredients into a shopping list, with smart units and pantry deduplication.
• Yours forever — your recipe book is portable, exportable, and not held hostage by an algorithm.
• Family-friendly — share recipes and menus with friends in-app.
• 14-day free trial, no credit card needed.

👉 Sign up with my referral link — you get the trial, and I get a tiny thank-you. Win/win.

✨ What this gives you

• 🤖 MCP server out of the box — point Claude Desktop, Claude Code, OpenClaw, or any MCP-compatible host at it and start talking to your recipe book.
• ⌨️ A CLI over the same tools — every MCP tool is also a plan-to-eat subcommand, with tables for humans and --json for scripts. One registry feeds both surfaces, so they can't drift.
• 🔐 Auto-auth & auto-recovery — set your credentials once, the server handles login, caches the cookie session to disk, and silently re-auths whenever Plan to Eat invalidates it.
• 🧰 36 tools, all the verbs that matter — full recipe CRUD, full meal-planner CRUD (add / move / duplicate / delete recipes, notes, and ingredient entries), reorder events in a slot, leftovers as a first-class workflow, freezer tracking, browse courses & cuisines & tags, count what's in your queue.
• 🗓️ Real planner control — view a week's plan with recipe titles pre-joined, schedule recipes on dates, attach prep notes, reschedule with one tool call, change servings, duplicate, search for duplicates, reorder same-slot events.
• 🛒 A shopping list you can actually work with — read it with the store and aisle each item is filed under, add items (Plan to Eat guesses the aisle and reuses the store you last picked), retitle or re-quantify a line, move items between stores, remove them, put them back.
• ❄️ Freezer tracking — after cooking, mark N portions as frozen with freeze_recipe_portions; check what's stashed with list_frozen_recipes; consume entries when you eat them (soft-delete, history preserved).
• 🍳 Real CRUD — including ingredient lists with proper units, directions, prep/cook times, nutrition, ratings, and tags.
• 📦 Tiny runtime — no Playwright, no headless browser, no native modules. Just fetch, the MCP SDK, and Zod. Boots in under a second.
• 🦺 100% TypeScript — fully typed Recipe, Ingredient, PlannerEvent shapes plus a generic _json<T> so your tools never have to guess what comes back.
• 📚 A library too — core/client.ts is a clean, plain-Node API client you can drop into any script.
• 🧠 Bundled Claude Code skill — .claude/skills/plan-to-eat/SKILL.md teaches any agent the common workflows and gotchas (the supper-vs-dinner alias, the description-vs-title mismatch, etc.) so it doesn't have to rediscover them.

Reverse-engineered from the live web app. There's no public Plan to Eat API, but the desktop site uses these same endpoints internally. Use at your own risk — they could change anything at any time.

🚀 Quick start

npx -y plan-to-eat-mcp          # run the MCP server
npm i -g plan-to-eat-mcp        # …or install both bins on PATH

Requires Node 18+ (for built-in fetch) and a Plan to Eat account.

The package ships two bins: plan-to-eat-mcp (the MCP server) and plan-to-eat (the CLI). To run the CLI through npx without installing, note that you have to select it explicitly, since the default bin is the server:

npx -y -p plan-to-eat-mcp plan-to-eat --help

git clone https://github.com/alex-zwingli/plan-to-eat-mcp.git
cd plan-to-eat-mcp
npm install
npm run build

Configuration

The CLI also reads a .env in the working directory (shell variables win). The MCP server does not — MCP hosts pass env explicitly, as shown below.

🤖 Use it with an agent

Any host that can launch a local stdio MCP server works. The server needs one command, two env vars, and nothing else — no ports, no OAuth, no daemon.

Claude Code, the easy way: install the plugin

The repo ships as a Claude Code plugin — MCP server and both skills in one step:

export PLAN_TO_EAT_USERNAME=you@example.com
export PLAN_TO_EAT_PASSWORD=hunter2

git clone https://github.com/alex-zwingli/plan-to-eat-mcp.git
cd plan-to-eat-mcp && npm install && npm run build

claude plugin marketplace add "$(pwd)"
claude plugin install plan-to-eat@plan-to-eat

That gives you:

Confirm with claude plugin details plan-to-eat@plan-to-eat and claude mcp list. The server reads PLAN_TO_EAT_USERNAME / PLAN_TO_EAT_PASSWORD from the environment Claude Code was launched with, so export them in your shell profile rather than committing them anywhere.

Claude Code, manually

claude mcp add plan-to-eat \
  --env PLAN_TO_EAT_USERNAME=you@example.com \
  --env PLAN_TO_EAT_PASSWORD=hunter2 \
  -- npx -y plan-to-eat-mcp

Add --scope user to make it available in every project instead of just this one. Check it connected with claude mcp list, or /mcp inside a session.

Claude Desktop

Edit claude_desktop_config.json — Settings → Developer → Edit Config, or:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "plan-to-eat": {
      "command": "npx",
      "args": ["-y", "plan-to-eat-mcp"],
      "env": {
        "PLAN_TO_EAT_USERNAME": "you@example.com",
        "PLAN_TO_EAT_PASSWORD": "hunter2"
      }
    }
  }
}

Restart Claude Desktop, and you're cooking.

Any other MCP host

Cursor, Windsurf, Zed, Cline, Continue, OpenClaw, VS Code's MCP support, and custom SDK clients all take the same three things. Point them at:

Most of them use the same mcpServers JSON block as Claude Desktop above — often in .cursor/mcp.json, .vscode/mcp.json, or the host's settings UI.

If you installed globally (npm i -g plan-to-eat-mcp), use the plan-to-eat-mcp bin as the command and drop the args entirely. If you built from a clone, the command is node with ["/absolute/path/to/plan-to-eat-mcp/dist/mcp/server.js"].

Absolute paths matter for the clone route. MCP hosts don't launch servers from your project directory, so a relative path will fail to resolve. The npx command above sidesteps this entirely.

Verify it works

PLAN_TO_EAT_USERNAME=you@example.com PLAN_TO_EAT_PASSWORD=hunter2 npx -y plan-to-eat-mcp

You should see [plan-to-eat] mcp server ready on stdio (36 tools) on stderr. That's the server waiting for a client — Ctrl-C out. If instead you get a credentials error, fix that before wiring up a host, where the failure is harder to see.

Teach your agent the workflows

Two skills ship in .claude/skills/, covering the common flows and the sharp edges (the supper-vs-dinner alias, the description-vs-title mismatch on note entries, checking for duplicates before scheduling):

• plan-to-eat — for agents calling the MCP tools.
• plan-to-eat-cli — for agents that have a shell but no MCP server. Same 30 capabilities, driven through subcommands, with --json for parsing.

The plugin install above registers both. Agents on other hosts can read them as plain context — point them at the files, or paste one into your system prompt.

Or install just the skills, into any of 18+ agents, with the skills.sh CLI:

npx skills add alex-zwingli/plan-to-eat-mcp

That copies both SKILL.md files into .agents/skills/ and symlinks them for Claude Code, Cursor, Codex, Copilot, Gemini CLI and the rest. Add --skill plan-to-eat-cli to take only one. Note that this installs the skills and not the server — each skill's setup section walks the agent through building the MCP server or CLI if it isn't already there.

For OpenClaw agents, both skills are on ClawHub:

clawhub install plan-to-eat        # MCP tools
clawhub install plan-to-eat-cli    # shell / CLI

Their frontmatter declares the credentials and binaries each one needs under metadata.openclaw, so ClawHub can check your environment at install time.

⌨️ Use it from the terminal

Every MCP tool is also a subcommand. Underscores become dashes; both spellings work.

npm i -g plan-to-eat-mcp
plan-to-eat --help

Or without installing — note the -p, since the package's default bin is the MCP server, not the CLI:

npx -y -p plan-to-eat-mcp plan-to-eat --help

$ plan-to-eat list-frozen-recipes
ID      RECIPE_ID  COUNT  SERVINGS  FROZEN_ON
──────  ─────────  ─────  ────────  ──────────
227854  44676380   6      1.0       2025-05-04
227908  17314561   3      1.0       2025-05-05

2 rows

$ plan-to-eat add-planner-note "Defrost chicken" 2026-05-05 dinner
id       67710999
date     2026-05-05
section  dinner

$ plan-to-eat get-counts --json | jq .frozen
10

• Positional or flagged — get-recipe 123 and get-recipe --id 123 are the same. <command> --help lists which arguments are positional.
• --json prints the raw payload instead of a table, for piping into jq.
• Arrays repeat the flag (--event_ids 11 --event_ids 22) or take JSON (--event_ids '[11,22]'). Object arguments take JSON: --ingredients '{"title":"bread","amount":"2"}'.
• Validation is the same Zod schema the MCP server uses, so a bad enum or a malformed date fails the same way in both surfaces.

🧰 The 36 tools

Each is an MCP tool and a CLI subcommand — add_planner_recipe the tool is plan-to-eat add-planner-recipe in the shell. Full reference with input schemas and return shapes: docs/TOOLS.md.

Recipes

Planner — read

Planner — write

Freezer

Shopping list

A line is addressed by its item_ids array, not a scalar id — Plan to Eat merges duplicate ingredients into one line that keeps every underlying row id.

Lookup tables & extras

📚 Use the client as a library

import { PlanToEat, type Recipe } from 'plan-to-eat-mcp';

const pte = new PlanToEat();
await pte.login(process.env.PLAN_TO_EAT_USERNAME!, process.env.PLAN_TO_EAT_PASSWORD!);
// Or restore: pte.importSession(JSON.parse(fs.readFileSync('session.json', 'utf8')));

const recipes = await pte.listRecipes();
const detail: Recipe = await pte.getRecipe(recipes[0].id);

const created = await pte.createRecipe({
  title: 'Grilled cheese',
  servings: 1,
  directions: 'Butter bread. Cheese. Pan. Flip.',
  ingredients: [
    { title: 'bread',  amount: '2', unit: 'slice', position: 1 },
    { title: 'cheese', amount: '1', unit: 'slice', position: 2 },
  ],
  tag_titles: 'easy,lunch',
});

await pte.updateRecipe(created.id, { rating: 5 });
await pte.deleteRecipe(created.id);

The client stashes credentials internally on first login so it can transparently re-authenticate if the cached cookies expire mid-session.

The tool registry is exported too, if you want to build your own adapter over the same 36 capabilities:

import { toolsByName, createSession } from 'plan-to-eat-mcp';

const { pte, ensure } = createSession();
await ensure();

const getRecipe = toolsByName.get('get_recipe')!;
console.log(getRecipe.description);          // same text the MCP host sees
const recipe = await getRecipe.run(pte, { id: 123 });

🧪 API reference (the parts that work)

All paths under https://app.plantoeat.com.

JSON endpoints (/api/v1/*)

Planner write endpoints (/planner/*)

A different style: form-encoded bodies, text/javascript (empty) responses. The server mutates state and the UI re-fetches separately. Required headers: X-CSRF-Token, X-Requested-With: XMLHttpRequest, Accept: text/javascript. Because the response body is empty, the client recovers any new event ID by diffing /api/v1/events before and after.

section values are breakfast, lunch, dinner, snacks. Note that the server may normalize dinner → supper based on the user's per-account preference; reads will reflect the canonical name.

Authentication

Cookie-based. The client supports two flows:

1. Login. POST /login with login[email] and login[password] plus the authenticity_token from the meta tag on GET /login. Server returns a 302 and sets cookies.
2. Reuse cookies. The "remember me" pair ptermid2 (user id) and ptermxt2 (long-lived token) is sufficient on its own — stash those and skip re-login until they're invalidated.

Write requests need an X-CSRF-Token header. The client grabs it from the <meta name="csrf-token"> tag on the /recipes page on demand.

HTTP Basic auth on /api/v1/... is not supported (returns 401).

Recipe payload shape

The single-recipe response has 65 fields. Writable on POST/PUT include: title, description, source, url, servings, yield, scaling, prep_time, cook_time, total_time (minutes), course_id, cuisine_id, main_ingredient_id, tag_titles (comma list), directions (free text), private, draft, rating, nutrition strings (calories, sodium, etc.), and ingredients.

Ingredients use Rails nested-attributes — the wire field is recipe_ingredients_attributes, not ingredients. The client and the create_recipe / update_recipe tools accept the friendlier name ingredients and translate. Each entry:

{ "title": "water", "amount": "1", "unit": "cup", "note": "", "position": 1 }

To delete an existing ingredient on update, include its id plus "_destroy": true. The server back-fills amount_float, metric_amount, metric_unit, and similar_titles.

📁 Layout

One package, three layers. The logic lives in core/; the MCP server and the CLI are thin adapters over the same registry.

src/
  core/
    client.ts     API client — no MCP, no CLI, no I/O beyond fetch
    tools.ts      the tool registry: name, description, Zod shape, handler
    session.ts    credentials + cookie-cache bootstrap, shared by both adapters
    version.ts    version read from package.json at runtime
  mcp/server.ts   registry -> MCP tools (stdio transport)
  cli/
    main.ts       registry -> subcommands, help, dispatch
    args.ts       argv -> Zod-validated arguments
    render.ts     results -> tables (or raw JSON under --json)
  scripts/        smoke tests and the docs coverage check
  index.ts        public entrypoint for library consumers
  server.ts       back-compat shim, see below

Adding a tool means adding one entry to src/core/tools.ts. It appears in the MCP server and the CLI at once, with the same name, schema, and description — they can't drift apart.

Other files worth knowing:

• docs/TOOLS.md — per-tool reference with input/output shapes and gotchas. Hand-written; npm run check:docs fails if it and the registry disagree about which tools exist.
• .claude/skills/plan-to-eat/SKILL.md — how to use the MCP tools well.
• .claude/skills/plan-to-eat-cli/SKILL.md — the same, for the CLI.
• .claude-plugin/ — Claude Code plugin and marketplace manifests.
• .mcp.json — the plugin's MCP server declaration.
• dist/ — emitted by npm run build.

Upgrading from ≤0.4? The server moved from dist/server.js to dist/mcp/server.js. The old path still works — it's a shim that loads the new one — so existing host configs and deployments keep running. New configs should use the new path or the plan-to-eat-mcp bin.

Playwright was used during reverse engineering and is kept as a devDependency for any future API-discovery work; the runtime depends only on fetch, the MCP SDK, and Zod.

Scripts

The verify* and test scripts hit your real account. They create and delete their own test data, but they are not a dry run — see CONTRIBUTING.md.

🤝 Contributing

Bug reports, new endpoints, and upstream-breakage fixes are all welcome — see CONTRIBUTING.md. The one thing to know up front: the test scripts run against a real Plan to Eat account and create real data (then clean it up). Use your own.

🙏 Support Plan to Eat

This whole project exists because Plan to Eat is great. If you find this useful, the best thing you can do is give Plan to Eat a try with my referral link. Free 14-day trial, no card needed.

License

MIT — see LICENSE. Do whatever you want with it; just don't blame me if Plan to Eat ships a breaking change.