午餐订单管理服务器

Relish MCP Server

MCP (Model Context Protocol) server for Relish by ezCater — manage corporate-subsidized lunch orders from your CLI or AI agents.

Usage

Once set up, just tell the agent:

order my lunches

That's it. The agent will check which days you haven't ordered on yet, pick food based on your saved preferences, and place orders. Run it daily or once a week — it only orders for days that are missing.

By default the agent shows you its picks and waits for a thumbs-up. If you'd rather it just handle everything, say "turn on auto-order" and it'll pick and order without asking.

Other things you can say:

Features

• Batch ordering — one command orders for every unordered day in the week
• Food preferences — remembers what you like and picks accordingly
• View schedule — see available restaurants, delivery times, and meal availability
• Browse menus — full menus with prices, categories, and dietary tags
• Place & cancel orders — end-to-end ordering and cancellation
• Subsidy tracking — stay within your company meal budget automatically
• 9AM backup auto-order (opt-in) — a cron job orders for today if you forgot; you stay in control until it fires
• Cookie persistence — MFA only needed once every 30 days

Setup

1. Install Chrome/Chromium

The server uses Selenium with a headless Chrome browser. You need Chrome or Chromium installed:

Fedora/RHEL:

sudo dnf install chromium chromedriver

Ubuntu/Debian:

sudo apt install chromium-browser chromium-chromedriver

macOS (Homebrew):

brew install --cask google-chrome

ChromeDriver is bundled automatically via selenium-manager in recent Selenium versions.

Verify it works:

chromium --version   # or google-chrome --version

2. Install Python dependencies

Requires Python 3.11+.

pip install -r requirements.txt

3. Run the server

With Cursor IDE (recommended):

The .cursor/mcp.json is already configured — just open this project in Cursor and the relish MCP server will be available.

Note: The MCP config runs python3 server.py. If you installed dependencies in a virtualenv, update .cursor/mcp.json to use your venv's Python path (e.g. ".venv/bin/python3"), or install globally.

Set RELISH_HEADLESS=0 in .cursor/mcp.json env to see the browser window (useful for debugging).

With Claude Code:

The project ships a .mcp.json, so just run claude inside this directory and approve the relish server when prompted. CLAUDE.md gives the agent the same guidance Cursor gets from .cursor/rules/. Everything — setup, ordering, the 9AM backup — works identically.

Standalone (stdio transport):

python server.py

4. First-time setup (agent-guided)

The agent handles all first-time configuration for you:

1. Credentials — the agent will ask for your Relish email and password and save them securely (.credentials, chmod 600, gitignored). You can also set them manually or via environment variables (RELISH_EMAIL, RELISH_PASSWORD).
2. MFA — on first login, a verification code is emailed to you. The agent will ask you to paste the 6-digit code. After that, cookies are saved and MFA is skipped for up to 30 days. (Only the Relish session cookies are saved, so if that session expires before the 30-day window you'll be asked for a code again.)
3. Food preferences — the agent will ask what cuisines you like/dislike and how you prefer to order, then save your preferences (.food_preferences.json, gitignored).

All of this happens conversationally — just start chatting with the agent and it will walk you through it.

Tools

Architecture

server.py              — MCP tool definitions (thin wrapper)
relish_browser.py      — All Selenium automation logic
relish_models.py       — Dataclasses (DaySchedule, MenuItem, Order, etc.)
.credentials           — Login credentials (chmod 600, gitignored)
.cookies.json          — Session cookies (gitignored)
.food_preferences.json — Dietary preferences (gitignored)

9AM backup auto-order (opt-in)

During setup (or anytime), the agent will ask — never assume — whether you want a weekday safety net:

"Want me to install a weekday 9:00 AM backup that orders for today if you haven't?"

Say yes and it calls setup_auto_order_cron, which installs a crontab entry running auto_order.py every weekday. The design keeps you in control and mistake-proof:

• You choose until it fires — order manually any morning and the backup sees your order and does nothing.
• Today only — it never orders for future days.
• Deterministic — the cron job is plain Python (no AI at runtime): preference keywords rank restaurants and items, and it picks the most food that fits your subsidy after tax. If it can't read the subsidy, it refuses to order rather than risk charging you.
• Transparent — ask the agent to run preview_auto_order to see exactly what it would pick today (without ordering), and auto_order_status for the run log (auto_order.log).
• Reversible — "turn off the backup" → remove_auto_order_cron.

If your login cookies expire, the backup logs a warning instead of ordering; a quick login through the agent fixes it. Requires crontab (Linux/macOS).

Food preferences

Stored in .food_preferences.json (gitignored). On first use, the agent will ask what you like and save your preferences. You can also edit the file directly:

{
  "yes": ["Mexican bowls", "Greek", "Mediterranean"],
  "no": ["Sushi"],
  "style": "Prefer bowls and plates over sandwiches",
  "notes": "Pick restaurants tagged 'Office favorite' when they match preferences",
  "auto_order": false
}

Set auto_order to true if you want the agent to pick and order without confirmation.

Update anytime via the set_food_preferences tool or by editing the file.