Food Near Me — MCP Server

Model Context Protocol server for AI-native restaurant discovery — three-tier search (verified → menu_indexed → discovered), Menu Protocol menus, and structured menu validation. Plug into Claude Desktop, Cursor, ChatGPT, or any MCP host in about 30 seconds.

Production endpoint: https://foodnear.me/mcp · 8 tools · 4 resources · 3 prompts · No API key (beta)

Quick start {#quick-start}

1. Add this to your MCP host config

Cursor — ~/.cursor/mcp.json (macOS/Linux) or %USERPROFILE%\.cursor\mcp.json (Windows)

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows)

{
  "mcpServers": {
    "foodnear-me": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://foodnear.me/mcp"]
    }
  }
}

2. Restart your MCP host

3. Try a prompt

“Find vegan Thai restaurants near Brooklyn Bridge and show me a menu for the top result.”

Your agent should call search_restaurants → get_menu (or get_restaurant first).

What you get

Tools (8)

Resources (4)

Prompts (3)

Configuration

Preview / local: Replace the URL with http://localhost:3000/mcp when running npm run dev in apps/web.

Operational tool filter: set FNM_MCP_ENABLED_TOOLS=search_restaurants,get_menu to expose only a comma-separated subset in tools/list and GET /mcp. Leave unset or * to expose all tools. This is for temporary degradation/context control, not privacy or auth.

Architecture

┌─────────────────────┐     POST /mcp (JSON-RPC)     ┌─────────────────────────┐
│  MCP host           │ ────────────────────────────▶│  apps/web/app/mcp       │
│  (Cursor / Claude)  │     GET /mcp (discovery)     │  Next.js route handler  │
└─────────────────────┘                              └────────────┬────────────┘
                                                                │
                                                                ▼
                                                   ┌─────────────────────────┐
                                                   │  Supabase + PostGIS     │
                                                   │  Menu Protocol (MP)     │
                                                   └─────────────────────────┘

Implementation: apps/web/app/mcp/route.ts · Flow runner: apps/web/lib/mcp/mcp-flow-runner.ts

Tool error contract

Failed tools/call responses include structured metadata in _meta.error:

Human-readable text is still in content[0].text for hosts that ignore _meta.

Verify

From repo root (with apps/web dev server running for localhost):

# Automated agent flows (14 flows when Supabase + seed configured; 11 without DB)
npm run test:mcp-flows

# Against production
npm run test:mcp-flows:http

# Discovery GETs + MCP tools/list count
npm run smoke:mcp

# Three-tier trust model copy parity (local files)
npm run check:discovery-copy

# Full deploy gate (13 checks + discovery copy on production URL)
npm run preflight -w web
# or: ./apps/web/scripts/deploy-preflight.sh https://foodnear.me

Production monitoring: GitHub Actions workflow MCP Production Smoke runs smoke:mcp daily and on manual dispatch (.github/workflows/mcp-smoke.yml).

Agent discovery

Scripted flows: apps/web/docs/example-agent-flows.md

Data trust model (three-tier search)

• search_restaurants returns verified → menu_indexed → discovered.
• Every result includes verification_status and menu_available. Call get_menu only when menu_available is true.
• Verified — owner-approved MP; authoritative for dietary/allergen claims.
• menu_indexed — automated/public MP menu; cite with caveat — not owner-verified.
• discovered — place only; do not cite menu items.
• Trust progression: discovered → menu_indexed → verified. See https://foodnear.me/attribution for data sources.

FAQ

Do I need an API key?
No for beta MCP access. Future paid tiers may use API keys or x402 (USDC on Base). See x402-prepaid-spec.md in your local docs/Food Near Me playbook.

Tools not showing after restart?
Confirm the config URL ends with /mcp. Restart the host completely. Run npm run smoke:mcp against your target base URL.

Empty search results?
Beta verified menus are seeded for specific metros (e.g. Williamsburg, NYC). 7 menu_indexed restaurants in Williamsburg have automated menus from website ingest. Discovered place listings cover many US metros — use coordinates in an imported region. Demo coords: 40.7128, -74.006. Run npm run db:seed -w web locally for verified test data.

Cursor vs Claude config path?
See Quick start above — each host uses a different JSON file; only the mcpServers block matters.

How is this different from DoorDash / Uber Eats APIs?
We expose owner-verified Menu Protocol data for agents — not scraped aggregator menus or ordering checkout.

Monorepo layout

This repository ships the MCP server inside the foodnear.me web app:

Business strategy and runbooks live in a separate local docs folder (not in this repo) — see your team's docs/Food Near Me playbook.

Development

npm install
cd apps/web && cp .env.example .env.local   # Supabase keys
npm run dev                                 # http://localhost:3000
npm run test:mcp-flows                      # POST localhost:3000/mcp

Operator: menu_indexed website ingest

Promote discovered → menu_indexed via free website/ordering-platform parsers (ChowNow API, order.online, Sauce, Squarespace, BentoBox, Toast, Playwright). Always dry-run first — headless is slow.

cd apps/web
npm run db:probe:menu-batch -- --headless --limit=10
npm run db:import:menu-indexed:website:headless:dry-run -- --limit=10
npm run db:import:menu-indexed:website:headless -- --limit=10   # live

No Uber Eats / DoorDash / Grubhub / RapidAPI scrapers. See apps/web/docs/example-agent-flows.md.

Links

• Website: https://foodnear.me
• GitHub: https://github.com/food-near-me/platform
• Menu Protocol spec: https://github.com/foodnearme/menu-protocol
• Support: https://foodnear.me/support · api@foodnear.me