Claude can explain, compare and turn restaurant numbers into text or tables as soon as it knows your data. There are two ways to get there: you upload export files, or you connect Claude via MCP directly to the system where your numbers live. This article shows both routes, ten questions for everyday use and the limits you should know about. Where these queries fit into a weekly routine is shown in the model case of a business with eight locations in the article on restaurant financial controlling.
Summary
- Without a connection, Claude works with whatever you paste or upload, for example a POS export as a CSV file.
- With a connection, Claude fetches the numbers itself. For Kiosa customers this works through the Kiosa AI plugin; for OktoPOS, through the vendor's own MCP interface.
- Claude only sees what the connection releases. With Kiosa, that is the data you are allowed to see in Kiosa, read-only by default.
- Check calculations and recommendations in your operation before you act on them.
What can Claude do with operating numbers?
Claude is a language model from Anthropic. It reads tables, summarises them, answers questions about them and phrases results so you can use them in a meeting. For you as an operator that means: instead of clicking through reports, you ask in plain language.
Typical tasks are comparing locations, explaining a deviation, summarising a week for the team or drafting an email to a supplier. What Claude cannot do is know numbers it has never seen. Without data it guesses, and guessed numbers are no use in controlling.
Without a plugin: upload files
The simplest route is the file route. You export a report from your POS or scheduling tool and give it to Claude. This works with any system that can export and needs no setup.
The limits show up in daily use:
- The file shows the state at the time of export. For current numbers you export again.
- Everything in the file goes to the assistant. A file has no roles.
- With several locations, you merge the files yourself.
- With large tables, calculation errors can occur. Spot-check totals and percentages against your system.
A detailed guide to this route is in the article on how to analyze restaurant POS data with ChatGPT. The advice there applies to Claude in exactly the same way.
With a plugin: connect Claude via MCP
MCP stands for Model Context Protocol. It is an open standard that lets an assistant such as Claude work with a business system. Instead of a file, Claude gets tools it uses to query numbers.
The Kiosa AI plugin uses this standard. Here is how:
- Your admin enables AI plugins in Kiosa under Settings in the MCP section and chooses the roles.
- In Claude you enter the address
https://mcp.kiosa.ai/mcpas a custom connector. Anthropic lists the steps in its help article on custom connectors. - You sign in to Kiosa and confirm access.
After that, Claude only sees what you are allowed to see in Kiosa, depending on your role and the modules you have booked. The plugin is read-only by default. Writing is a separate admin switch that you must also allow when connecting. All the details are on the Kiosa AI plugin page.
Ten questions you can ask Claude
These questions match the data the Kiosa AI plugin can retrieve. Which ones you get answered depends on your role and your modules.
- How were sales yesterday across my locations?
- Which location is furthest below last year this week?
- Which items are selling best and worst this week?
- What was the labor cost percentage last month, and how does it differ by location?
- Show me the rota for this week.
- Which items have seen the biggest recent increases in purchase prices?
- How are discounts and voids developing by location?
- Which invoices are open, and from which suppliers?
- What do the latest reviews say, and what stands out?
- Which reservations are there for the weekend?
Good questions name a period and a comparison. So do not just ask about sales; ask about last week's sales against the week before or against last year.
OktoPOS offers its own MCP interface
OktoPOS also describes an MCP interface for the OktoPOS Manager. According to the vendor page, an MCP-capable agent such as Claude or ChatGPT can query sales and check data, such as sales, number of checks and average check per period and branch. OktoPOS recommends setting up the agent as a separate user with its own role, using read access for reporting and having a person approve consequential steps such as price changes.
If you use OktoPOS exclusively and want to ask POS questions, that is a direct route. The Kiosa AI plugin is worth it when your questions touch several data sources, for example POS sales together with labor cost percentage, food cost or invoices across all locations. How Kiosa uses OktoPOS data is described on the OktoPOS integration page.
Limits: what to keep in mind
- Claude does not calculate without errors. Check important totals and percentages against your system, especially with files.
- Answers are drafts for decisions. You need to verify causes and recommendations in your operation. A drop can be down to weather, roadworks or staffing, none of which shows up in a number.
- The data sets the limit. If a module or permission is missing, Claude does not get the number.
- Data protection remains your job. Staff and customer data fall under data protection law (in the EU, the GDPR). On the file route, you decide what to upload. With the plugin, roles and admin approval decide.
- Read first, write later. Keep write permissions switched off until you really need them.
For a sober look at what AI does and does not do in restaurant controlling, see the article on AI in restaurants.
What it costs
The Kiosa AI plugin requires a Kiosa account. One View costs €99 per location per month, with no setup fee and a 12-month minimum term. The modules and all prices are on the pricing page. The cost of Claude itself depends on the plan you choose with Anthropic. If you run several locations, the quick-service and multi-unit restaurants page also helps you get oriented.
From analysis to action
What this is about: Claude with your operating figures: file or MCP plugin, ten example questions and the limits.
Put it into practice:
More on Controlling and finance
- Restaurant Financial Controlling: What to Measure Daily, Weekly and Monthly
- Analyze Restaurant POS Data with ChatGPT: CSV vs. Plugin
- AI in Restaurants 2026: What Actually Works
FAQ
Can Claude analyse my POS data?
Yes, in two ways. You can upload an export as a file, or you can connect Claude via MCP to a system that holds your numbers. For Kiosa customers this works through the Kiosa AI plugin; for OktoPOS, through the vendor's MCP interface.
Is Claude suitable for restaurants?
For analysis, summaries and text drafts, yes, as long as Claude knows your real numbers. It does not replace a controlling system, but it makes a system's numbers accessible by question. Check important results before you base decisions on them.
What is MCP and do I need it?
MCP is an open standard that lets assistants such as Claude work with business systems. You do not need it for one-off questions about a file. It pays off when you regularly ask questions about current numbers for several locations.
Does Claude see all my data in Kiosa?
No. Claude only sees what you are allowed to see in Kiosa, based on your role and your business's modules. By default the plugin only reads. The admin can disconnect connections at any time.
Can Claude carry out tasks in my business?
Only if the admin enables writing in Kiosa and you allow it when connecting. Claude can then, for example, comment on tasks in the workspace or create a draft reply to a review, which waits in Kiosa for your approval.