
MCP (Model Context Protocol) is a standard plug that lets an AI assistant, such as ChatGPT, Claude or Mistral, connect to a business's tools. For a restaurant or a trade business, it means asking questions in the AI you already use, about your own data rather than the internet at large. The plug guarantees nothing on its own: what protects you are the rules of the connected software, starting with rights identical to the person's and actions that wait for their confirmation.
Monday, 3:40 p.m., restaurant closed
Véronique runs a forty-cover restaurant. Mondays, it is closed. She sits at the counter with her laptop and the notebook where she keeps her supplier orders. For months she has used an AI assistant to draft emails and polish her menu wording. This time she asks it something else: "Which closing checklists were not done last week?"
The assistant answers in three lines: two evenings with no closing checklist at the dish station, Tuesday and Saturday, and one checklist done on Friday but with no photo attached. It says where each piece of information comes from. Véronique goes on: "Create a task for Damien: go over closing with the evening team again, before Friday." The assistant creates nothing yet. It shows a summary (the task, the person, the due date) and waits for her to confirm in a new message.
Then, almost absent-mindedly, Véronique types: "Note that Charlotte is pregnant, she'll be less available." The assistant refuses to record that and offers to log a dated fact instead, such as declared unavailability on specific dates.
There is nothing exceptional about this scene: a general-purpose AI, connected through a standard plug to the tool where the restaurant's information lives, and rules that hold.
(Scene and characters are fictional.)
What is MCP, in plain words?
A standard plug between an AI and your tools
Think of a wall socket. The toaster knows nothing about the power station, and the power station knows nothing about the toaster. They work together because the socket is the same everywhere.
MCP plays that role between an AI assistant and business software. On one side, the assistant you already use. On the other, the tool that holds your schedules, tasks, checklists or sales. MCP sets out a shared way for them to talk: the assistant asks, the tool answers with what it agrees to show, and it can suggest actions.
Before this standard, every link between an AI and a piece of software was a separate build, redone for each assistant. With a shared plug, a software vendor builds it once and several assistants can connect.
Where it comes from, and who has adopted it
The Model Context Protocol was released by Anthropic, the maker of Claude, on November 25, 2024, as an open standard (Anthropic announcement, 2024). Open means any vendor can use it, with no license.
Others then picked it up. OpenAI announced support for it in March 2025. In September 2025, Mistral AI opened MCP-based connectors in Le Chat, with the option to add more (Mistral AI announcement, September 2, 2025). In December 2025, Anthropic handed the protocol to the Agentic AI Foundation, hosted by the Linux Foundation, co-founded with Block and OpenAI, and backed by Google and Microsoft among others (Linux Foundation press release, December 9, 2025).
For a restaurant owner, the takeaway is simple: the connection no longer depends on a single AI provider.
What MCP is not
MCP has no intelligence of its own, and nothing gets installed in your kitchen: it simply describes how two programs talk to each other.
Above all, MCP does not decide what the assistant may see or do. The connected software sets those limits. Two tools can use the same plug and offer very different safeguards. That is why the rest of this article is mostly about rules, and very little about technology.
Why it matters for a restaurant or a trade business
Ask your questions where you already work
Many small-business owners already keep an AI assistant open on their laptop or phone. They use it for emails, job descriptions, messages to customers. MCP lets them ask about the business there too: "What have my managers not dealt with this week?", "What waste was logged at the downtown site, and why?".
You do not switch tools to think. You keep the one you know, and it goes and fetches the information. We cover this use in asking your restaurant questions from ChatGPT, Claude or Mistral.
About your own data, not the internet
A general AI knows a lot about restaurants in general. It knows nothing about yours: your standards, your stations, what was done last night. Unconnected, it fills gaps with averages, or makes things up. Connected to the tool where your facts live, it can answer from what was actually recorded, and say when data is missing. More on this in why a general AI does not know your restaurant.
Analyzing and acting are two separate uses
Once connected, the AI can do two very different things. It can analyze: line up facts, sum up a week, spot what keeps coming back. It can also act: create a task, change a schedule, send an instruction. The first changes nothing in the business. The second does. The rules cannot be the same, as AI that analyses and AI that acts: two different jobs explains.
What to demand before connecting your AI
Since the plug itself guarantees nothing, the connected software is what you hold to account. Here is what a restaurant owner or tradesperson should demand, whatever the vendor.
- Rights identical to the person's. An assistant connected through a line cook's account should only see what that line cook already sees. Not the other sites' schedules, not colleagues' one-to-ones. A plug should open no new doors. See give your AI the same rights as you, no more.
- Two-step actions. The AI prepares and sums up the effect, the person confirms. A misunderstood sentence should trigger nothing. Details in AI proposes, a human confirms: the two-step rule.
- A log. Every confirmed action should be recorded: who, when, what. The day a task disappears or a shift moves, you need to be able to find out what happened.
- Clear exclusions. Some topics carry too much weight to go through a chat: payroll, invoicing, permanent deletion and legal acts. They should stay out of the AI's reach, even with confirmation.
- Refusal of sensitive data. An employee's health, a pregnancy, a disability, their religion or opinions are special-category data under Article 9 of the GDPR (the EU data protection regulation). A serious tool refuses to record them, however the request is worded. We cover this in MCP and sensitive data: what AI must never record about an employee.
- An honest answer when information is missing. The tool should say "no data" rather than let the AI fill the gap.
Those demands at a glance
| Demand | Good answer from the vendor | Answer that should worry you |
|---|---|---|
| Rights | The AI sees and does what the person already sees and does | "It has access to everything, it's easier" |
| Actions | Prepared, summed up, confirmed by the person | "It acts directly, to save time" |
| Log | Who, when, what, available to review | "We don't keep track of actions" |
| Exclusions | Payroll, invoicing, permanent deletion and legal acts out of reach | "Anything goes as long as you confirm" |
| Sensitive data | Refused, with a dated, observed fact offered instead | "It's up to you to be careful" |
| Missing data | Flagged as missing | "The AI always finds an answer" |
The risks, without the drama
Giving too many rights
The most common risk has less to do with hacking than with convenience. To make it "work straight away", people connect the assistant with the owner's account, which sees everything, then share it with a manager. The manager now sees what they should not. The rule is simple: everyone connects their assistant with their own account, and the assistant inherits their rights, not someone else's.
Data sent to the AI provider
When you ask your assistant a question, the answer your software sends back goes through that assistant's provider. That is how it works: the AI has to read the information to answer you. So you need to know what leaves, to whom, and under what terms (free or business plan, terms of use, where it is processed). Good software limits how much can leave at once. And what it does not expose to the AI does not leave.
Actions without confirmation
AI sometimes gets it wrong. "Delete Damien's shift" can hit the wrong day, or the wrong Damien. If the action runs with no summary, the mistake is made before anyone sees it. Two-step confirmation turns a misreading into a draft you simply correct.
Smooth answers with no source
A well-written answer inspires trust, even when it rests on nothing. Connected to your data, the AI should be able to show where each fact comes from, and tell apart what was recorded, declared, proven or inferred. Otherwise, the plug only makes invention more convincing.
Common mistakes
- Connecting the assistant with an admin account "to test", then forgetting to change it.
- Judging the tool on how well it writes rather than on how accurate its facts are.
- Never testing a forbidden request. Ask the AI to score an employee or record health information: its answer will tell you more than a brochure.
- Confusing "the AI can read" with "the AI can act". They are two separate decisions, to take one after the other.
- Forgetting leavers. When an employee leaves, their access must be cut, including the assistant they had connected.
- Thinking the standard protects you on its own. Two products can use the same plug with opposite rules.
How Tsuno applies these rules
Tsuno offers this connection: ChatGPT, Claude, Mistral or another compatible assistant can query Tsuno. The Integrations page lists it among connected tools, and the Trust and data page sets out the rules around it. Here they are, as we publish them.
The assistant only sees and does what the person can already see and do in Tsuno. Reads are capped, so the connection cannot be used to pull out your data wholesale. Every action follows the same path, in Tsuno or from the assistant: Tsuno prepares, sums up the effect, the person reads the summary and confirms in a new message. Access is checked again at the moment of confirming: if a right was removed in the meantime, the action does not go through. The log keeps who, when and what, with the state before and after.
Some actions stay out of the assistant's reach, even when confirmed: payroll, invoicing, permanent deletion and legal acts. Tsuno refuses to record an employee's health, pregnancy, disability, religion or opinions, and offers to log dated, observed facts instead. It exposes no ranking of people to the assistant, and the instructions it gives the assistant ask for dated facts, never a judgment about someone.
On data, the app and its data are hosted in Germany, in the EU. When you use your own assistant, what it reads to answer you is processed by its provider, under your terms with them: that is a point to check on your side, and we would rather say so. When a right is removed or an employee leaves, their open sessions are closed.
What we do not claim: the AI only knows what is recorded in Tsuno. If the checklists are not done, it will say so, and that is useful.
San and your usual assistant: the voice at the station, the AI at the desk
Tsuno San is the intelligent voice assistant for frontline teams, connected to the same Tsuno environment: procedures, tasks, checklists, routine checks, workplace information. It relies on the same memory and the same rights as the assistant connected through MCP. The difference is about when and where.
At the station, in the kitchen or the stockroom, hands are busy. You need a short answer, out loud: how to clean the slicer, the next task, the check that is due. That is San's job. San Voice, the voice interaction, is available in demo; San Audio (hands-free) and San 360 (a broader vision of operational support) are under exploration and testing.
At the desk, in front of a screen, there is time to analyze, compare, prepare a meeting. That is where your usual assistant belongs. One memory, several ways to talk: we develop the idea in San and MCP: one memory, many ways to talk, and the difference between a general voice and a work companion in general voice assistant or companion that knows your workplace. The Tsuno San page presents the voice assistant.
| Situation | The right tool | Why |
|---|---|---|
| Prep cook, hands in the tub | San, by voice | A short answer without stopping the task |
| Manager on Monday, restaurant closed | Her usual AI assistant | Time to analyze and review |
| Manager preparing a one-to-one | Their AI assistant, or Tsuno directly | Dated facts to gather, not a judgment |
| Dishwasher reporting a leak | San, by voice | The report is prepared, he confirms |
What to measure
- The share of answers that say where the information comes from, out of ten real questions put to your connected assistant.
- The number of actions prepared by the AI and corrected before confirmation. A high number mostly shows that confirmation is earning its keep.
- Time spent, on Monday or at the end of the week, pulling together information from your sites, before and after connecting.
Before you start
Connecting an AI to your business is like preparing for a new hire: you decide what they see, what they can do alone, what they need signed off. We gathered those questions in 12 questions to ask before connecting an AI to your business, with good answers and warning signs.
Every article on this topic
This guide is the entry point to the topic. The 8 articles that go deeper are grouped below by use.
Understanding the idea
- Why a general AI does not know your restaurant: why an AI that knows food safety rules knows nothing about your fridge 2 until it is connected to your records.
- San and MCP: one memory, many ways to talk: how voice, screen and your AI assistant read the same memory under the same rules.
- AI at work glossary for restaurant owners: 38 terms explained in kitchen language: the 38 terms of AI at work (agent, MCP, hallucination, DPIA) explained with a kitchen example.
The rules that protect you
- Give your AI the same rights as you, no more: why the AI must see and do exactly what the person can already see and do.
- AI proposes, a human confirms: the two-step rule: how an AI prepares and summarises, a person reviews and confirms, and the pitfalls of that rule.
- AI that analyses and AI that acts: two different jobs: why reading and cross-checking are not governed like creating, changing or sending.
Putting it into practice
- Asking your restaurant questions from ChatGPT, Claude or Mistral: what your usual assistant can answer about your business once connected to Tsuno.
- 12 questions to ask before connecting an AI to your business: the 12 questions to put to a vendor, with the answers to insist on.
The four other reference guides
- Tsuno San: the work companion for frontline teams: what the team assistant is, what it does and its safeguards.
- AI for frontline managers: analyse, advise, reassure: the four useful roles of AI for a frontline manager.
- The future of work in hands-on jobs: what will change in hands-on jobs and what will stay human.
- AI, voice and privacy at work: the guide: the legal and design rules for an acceptable voice AI at work.
For the assistant's written limits, also read What San never does, and for the difference between an AI that chats and a tool built for the job, Conversational assistant or job assistant: where is the line?.
Key takeaways
MCP is a standard plug that lets you ask questions in the AI you already use, about your own data. Protection comes from the connected software, and it is up to you to check it: the same rights as the person, confirmed actions, a log, exclusions and refusal of sensitive data.
Frequently asked questions
What is MCP, in plain terms?
MCP (Model Context Protocol) is a standard plug that lets an AI assistant such as ChatGPT, Claude or Mistral connect to a business's tools to read information and prepare actions, without custom development for each assistant.
Who created the Model Context Protocol?
Anthropic released it as an open standard on November 25, 2024. Other players adopted it afterwards, including OpenAI and Mistral AI, and in December 2025 it was handed to the Agentic AI Foundation, hosted by the Linux Foundation.
Is it risky to connect my AI assistant to my business software?
The risk depends on the rules set by the software, not on the plug itself. Demand that the assistant has the same rights as the person, that every action waits for confirmation, that a log keeps track, and that payroll, invoicing, permanent deletion and legal acts stay out of reach.
Can my AI assistant see everything in my business once it is connected?
It should only see what you already see in the tool, through your own account. With Tsuno, the assistant only sees and does what you can already see and do, and reads are capped to prevent bulk extraction.
Do I need to be a developer to use MCP?
Usually not. The software vendor builds the plug. The user adds it to their AI assistant and uses it by asking questions in everyday language.
What is the difference between San and my AI assistant connected to Tsuno?
They rely on the same workplace memory and the same rights. San is the voice at the workstation, designed for teams in the kitchen or the stockroom; your usual assistant is more for the office, to analyze, prepare and review.