FoodGen

An MCP server that turns a dinner request into a grocery list + recipe on a mobile-friendly page — ready to be saved in Apple Notes.

“Dinner for 3 people (2 adults and 1 child), quick before soccer practice, no fish.”

Claude finds the dish, calls create_dinner_plan, and responds with a link like https://foodgen.instantoffr.com/n/Hxe7gVY_YYqRi2CCeZnV5g. On mobile:

• Grocery list grouped by store department, with check-off in the store (remembered on the device)

• Recipe with ingredients and steps below

• “Add to Notes” → share sheet → Notes saves everything as one note with bullet points

• …/ n/<id>.txt gives plain text, …/ n/<id>.json gives raw data

Architecture

• Node 24 + Express, MCP over Streamable HTTP at /mcp (stateless — no sessions, everything resides in SQLite)

• SQLite (node:sqlite, no native dependencies) in a Docker volume

• The model writes the reci pe; the server validates (zod), stores, and publishes it

• Plan URLs are 128-bit random — not guessable, not indexed; that is the entire access control

MCP surfaces

Related MCP server: Recipe Research MCP Server

Running

cp .env.example .env      # sjekk PUBLIC_URL
docker compose up -d --build

TLS and proxy are handled by Nginx Proxy Manager, which is already running and configured for port 2400. One thing is worth checking on the proxy host: MCP responses stream as SSE, so buffering and a 60-second read timeout can cut off the stream. Turn on Websockets Support, and optionally add this in the Advanced tab:

location /mcp {
    proxy_pass http://$server:$port;
    proxy_http_version 1.1;
    proxy_set_header Host $host;
    proxy_set_header Connection "";
    proxy_buffering off;
    proxy_cache off;
    proxy_read_timeout 3600s;
}

Local development without Docker:

PUBLIC_URL=http://localhost:2400 npm run dev

Connect to Claude and ChatGPT

• The Claude app / claude.ai / Cowork: Settings → Connectors → Add custom connector → https://foodgen.instantoffr.com/mcp (no authentication)

• Claude Code: claude mcp add --transport http foodgen https://foodgen.instantoffr.com/mcp

• ChatGPT: Settings → Connectors → add a custom MCP connector with the same address, no authentication. The server exposes search/fetch according to the OpenAI contract, so it works like a regular connector (search/lookup in stored dinners and deep research); with Developer mode on, ChatGPT can also use create_dinner_plan and create new dinners.

The front page at https://foodgen.instantoffr.com shows the same instructions.

Test

docker compose up -d --build   # eller npm start
node test/e2e.mjs              # full MCP + HTTP-runde, 26 sjekker

Environment variables