Insights MCP server

Connect ChatGPT, Claude, Copilot, or Gemini to your Fooodo Insights tenant via the remote Model Context Protocol server.

Fooodo Insights ships a per-tenant remote Model Context Protocol server. It exposes your organisation's operational analytics and a small set of bounded write actions to any AI client that speaks MCP — without API keys, custom integrations, or copy-paste.

Who can use it? Any Fooodo Insights customer with admin, analyst, or viewer access to a tenant that has the mcp_enabled feature flag turned on. The flag is dark by default for every org; contact support@fooodo.com to request enablement and receive your tenant's MCP URL.

Two MCP servers, one product family. This page covers the Insights MCP server, which exposes per-tenant operational analytics over OAuth 2.1. A separate marketing-handbook MCP server runs at /api/mcp on fooodo.com and is public — agents use it to learn what Fooodo is and to submit lead requests. The two surfaces are deliberately separate: handbook content vs. customer data.

Why MCP

MCP is the open standard AI clients use to talk to remote data and tools — the agent's equivalent of OAuth + REST. A single MCP server can be added once and reused by ChatGPT, Claude, Copilot, Gemini, Cursor, Windsurf, Zed, and the growing list of compliant clients.

For Fooodo Insights customers this means:

No vendor lock-in. Your analyst can ask the same question from ChatGPT today and Gemini tomorrow without re-configuring anything.
Bounded, audited actions. Write tools (e.g. trigger data import, acknowledge alert) require the insights:write scope and an admin or analyst role. Every write is recorded in the audit log with source: "mcp".
Tenant isolation. Tools are scoped to the user's organisation at the database level; cross-tenant access is impossible.

Connect from your AI

The Insights MCP server speaks OAuth 2.1 with Dynamic Client Registration, so every client below uses the same one-click flow: paste your tenant URL, click Connect, log in with your Fooodo Insights account, approve the scopes — done.

The snippets below use <INSIGHTS_MCP_URL> as a placeholder. Replace it with the URL Fooodo support sends you after enabling MCP for your tenant.

ChatGPT (Plus / Pro / Enterprise)

Open Settings → Connectors → Add custom connector (Developer Mode must be enabled).
Server URL: <INSIGHTS_MCP_URL>.
Approve the consent page when prompted.

Claude.ai (Pro and above)

Open Settings → Connectors → Add custom connector.
Name: Fooodo Insights. Server URL: <INSIGHTS_MCP_URL>.
Approve the consent page.

GitHub Copilot (VS Code 1.101+, JetBrains, Visual Studio)

Add to your workspace's .vscode/mcp.json:

{
  "servers": {
    "fooodo-insights": {
      "type": "http",
      "url": "<INSIGHTS_MCP_URL>"
    }
  }
}

If you're on Copilot Business/Enterprise, your admin must enable the MCP servers in Copilot policy first.

Gemini CLI

Add to your Gemini CLI configuration:

{
  "mcpServers": {
    "fooodo-insights": {
      "url": "<INSIGHTS_MCP_URL>",
      "oauth": true
    }
  }
}

Gemini Enterprise

In the Google Cloud console, register the server as a custom MCP data store:

Field	Value
Server URL	`<INSIGHTS_MCP_URL>`
Authentication	OAuth 2.0 (Dynamic Client Registration)
Required scopes	`insights:read`, `insights:write`

What the agent can do

The server exposes tools in three tiers. Clients should pick the lowest tier that answers the question — it's cheaper, faster, and deterministic.

Tier 1 — Data tools (fast, deterministic, read-only)

Tool	Returns
`get_period_summary`	Aggregated metrics for a date range — sales, COGS, EBIT, gross margin, top campaigns
`get_weekly_sales_data`	Week-by-week sales with active campaigns and cost data
`get_channel_stats`	Per-marketing-channel performance metrics — campaigns and acquisition channels, not the dine-in/delivery split (for that, use `get_channel_breakdown`)
`get_daily_performance`	Day-by-day sales by channel and restaurant
`get_product_performance`	Product/menu-item ranking by sales and quantity, with category breakdown
`get_cfo_pnl_data`	P&L breakdown — revenue, COGS, labor, OpEx, EBIT — for the CFO dashboard view
`get_restaurant_performance`	Per-restaurant / per-location sales breakdown for a date range, with each location's share of the chain total
`get_channel_breakdown`	Sales-channel split (dine-in / delivery / Wolt / Bolt / takeaway) with € and % share — the sales-side counterpart; marketing channels live in `get_channel_stats`
`get_check_metrics`	Average check (€, net of VAT), dishes per check, and check counts for a date range — with per-order-type and per-restaurant splits
`get_table_times`	Average table-closing time (minutes) and order counts — a service-speed signal — with optional order-type and restaurant splits
`get_waiter_times`	Per-waiter average table-closing times for a single restaurant (`restaurant_id` required)
`get_guest_feedback`	Guest satisfaction — NPS survey scores (food / service / restaurant) plus the latest Google review ratings per restaurant
`get_cash_position`	Current cash balance, weekly burn rate, runway, and per-entity balances
`get_goals_vs_actuals`	KPI targets (average check, dishes per check) and daily sales budget vs. actual, with order-count attainment
`get_monthly_sales_plan`	Forward-looking monthly sales plan — revenue and order-count budget per channel and restaurant, including future months
`get_marketing_spend`	Monthly marketing spend (from the P&L) and the campaign calendar for the range
`list_restaurants`	Restaurant directory (id, code, name, city) — resolves a name or code to the `restaurant_id` the other data tools filter on
`get_menu_sales`	Year-over-year dish-sales breakdown by restaurant or city (lunch-only or general), with an optional channel filter

Tier 2 — Specialist agents (single-domain LLM analysis)

Tool	Returns
`consult_financial_analyst`	Financial-only "why" analysis grounded in the period's data
`consult_marketing_strategist`	Marketing-only causal-uplift and campaign-attribution analysis
`consult_market_research`	Market context — seasonality, holidays, consumer-behaviour trends

Specialist answers are computed, not improvised: multi-step arithmetic is performed by a deterministic server-side calculator, not LLM mental math, and every figure carries a confidence label and a source citation. For profit-impact questions, fetch the P&L baseline with get_cfo_pnl_data first and pass the relevant figures (baseline EBIT, costs) into consult_financial_analyst — the specialist's own tools cover sales data only.

Tier 3 — Orchestrator (cross-domain synthesis)

Tool	Returns
`ask_orchestrator`	Routes a cross-domain question across all specialist agents and synthesises one answer

Search & fetch (ChatGPT Deep Research, generic clients)

Tool	Returns
`search`	List of insights and agent-action history matching a query — `{id, title, url, snippet}`
`fetch`	Full content of a document by id from `search` — `{id, title, text, url, metadata}`

Write tools (require `insights:write` + `admin`/`analyst` role)

Tool	Does
`trigger_data_import`	Triggers a configured scheduled file fetch immediately (async — returns a `job_id`)
`acknowledge_alert`	Marks an alert as acknowledged
`resolve_alert`	Resolves an alert
`generate_report`	Generates a business report — executive summary, campaign performance, channel analysis — for a date range; async (returns a `job_id`), optionally emailed
`regenerate_insights`	Re-runs the insights generator for the org (async)
`get_job_status`	Polls an async job's status

Prompts

weekly_business_review, margin_diagnosis — slash-command prompts for ChatGPT Apps and other prompt-aware clients.

Security and limits

OAuth 2.1 + PKCE, public clients, rotating refresh tokens. Access tokens are HS256 JWTs valid for one hour. Refresh tokens are valid for 60 days, stored as SHA-256 hashes, single-use on rotation.
Per-organisation gating via the mcp_enabled feature flag — dark by default for every org. Enablement is a support ticket, not a self-serve toggle, so customers can stage rollout.
Per-tier rate limits per user. Data tools 60/min; specialist and orchestrator 10 per 5 min; write tools 5 per 5 min.
Revocation. Users can revoke any client at any time from their account settings (/account/connected-apps on the Insights side). Admins can revoke org-wide.
Audit trail. Every write tool call is recorded in the audit log with source: "mcp", including the client name, the user, and the input payload.
Tenant isolation at the database layer. Every query is scoped to the user's organisation before it leaves the application; the server cannot return cross-tenant data.

For agent developers

Discovery URL: <INSIGHTS_MCP_URL>/.well-known/oauth-authorization-server (and /mcp suffixed, per RFC 8414 §3.1).
Protected-resource metadata: <INSIGHTS_MCP_URL>/.well-known/oauth-protected-resource/mcp (per RFC 9728).
Dynamic Client Registration: POST <INSIGHTS_MCP_URL>/register — open registration, but every authorisation still requires interactive consent.
Scopes: insights:read (default) and insights:write (for write tools).
Transport: Streamable HTTP, MCP protocol version 2025-06-18.

Standards followed: MCP 2025-06-18, RFC 6749 (OAuth 2.0), RFC 7636 (PKCE), RFC 7591 (Dynamic Client Registration), RFC 8414 (Authorization Server Metadata), RFC 9728 (Protected Resource Metadata).

Insights MCP server

On this page