menu-MENAmenu-MENA
HomePricingBlogContact UsGet Started
The menu-MENA API access screen showing the MCP access section: a generated AI-agent token valid for 24 or 48 hours, with a Regenerate button and the MCP endpoint URL.

Connect an AI Agent to Your Menu: The MCP Server

Platform & Tech

By menu-MENA Team

Published on September 7, 2026

Most people who connect something external to their menu use the public API: four read-only endpoints for pulling data into a POS integration or a reporting tool. That covers reading. It doesn't cover doing. The MCP server is the other half, it lets an AI agent, Claude, ChatGPT, or anything else that speaks the Model Context Protocol, actually take actions against your account, the same actions you'd take yourself in the admin panel.

A token, scoped to your account and built to expire

Connecting is a token, not a login flow, generated from API access in the admin panel, next to your regular API key. You pick how long it lasts, 24 or 48 hours, rather than getting a permanent credential to keep track of. The server resolves that token to a specific tenant on every request, and every tool call it handles after that is scoped to that tenant and branch. Regenerating kills the old token immediately, so cutting off access an agent shouldn't have anymore is one click, not a support ticket.

What's actually on offer

The tools registered on the server aren't a demo set, they mirror real sections of the admin panel: branches, coupons, customers, domains, menu items and categories, modifier groups, orders, PDF menu handling, reservations, account settings, and templates. Ask an agent to draft a weekend coupon, pull up a customer's order history, or update your restaurant's phone number, and it's calling the same underlying actions a staff member would trigger by clicking through the panel.

A diagnostics tool built for one specific problem

One tool exists purely to catch mistakes: a diagnostics pass over your catalog that flags issues by name, and the one worth calling out is the bilingual gap check. An item with a name typed in English but never given an Arabic one, or the reverse, is exactly the kind of thing that survives unnoticed in a catalog with a few hundred items. An agent can run that check on request instead of someone scrolling the whole list by eye.

Every action is still logged

Actions taken through the server don't bypass your audit log. Tool calls route through the same audit system the rest of the admin panel uses, so a change an agent makes shows up in your history attributed to that action, not as an edit with no explanation behind it. If something changes, you can tell whether a person clicked it or an agent called it.

Built for agents, not for browsers

The server handles its own CORS and authentication before a tool call ever runs, including rate limiting on the authentication step itself, so it's built to sit behind an AI client making programmatic calls rather than a browser tab someone's poking at directly. If you're building your own integration by hand, the public API is still the simpler starting point. If you want an AI agent to actually run parts of your menu management for you, this is what it talks to.

Frequently Asked Questions

The public API is four read-only REST endpoints, built for pulling your menu, orders, or reservations into another system. The MCP server goes further: it exposes real actions, adding a modifier group, updating restaurant info, generating a domain, so an AI agent that supports MCP can actually do things, not just read them.

Not to use it. If your AI client of choice already speaks MCP, you point it at the server with your tenant's token and start asking it to do things. Writing new tools or wiring it into a custom app is developer work, using the tools that already exist isn't.

The registered tools cover branches, coupons, customers, domains, menu items and categories, modifier groups, orders, PDF menu handling, reservations, settings, and templates. In practice that means an agent can be asked to draft a new coupon, check what's in a customer's order history, or update your restaurant's contact info, the same ground the admin panel covers.

There's a dedicated diagnostics tool built specifically for this. One thing it checks for by name is bilingual gaps, an item with an English name but no Arabic one, or the other way around, which is easy to miss scrolling through a long catalog by hand and exactly the kind of thing worth an agent double-checking.

No, and that's deliberate. From API access in the admin panel, you generate a token valid for either 24 or 48 hours, not a permanent credential sitting in a settings page. Regenerating invalidates the previous token immediately, so if you ever need to cut off an agent's access, that's the button.

No. The token you connect with resolves to your tenant specifically, and every tool call is scoped to that tenant and branch. There isn't a mode where an agent can reach across accounts, the server has no path to it.

Yes. Tool calls route through the same audit system as everything else in the admin panel, so a change made by an agent shows up in your history with an actor attached to it, not as an unexplained edit.