How to Connect AI Agents to Your Business Tools: A Guide for Restaurant Owners
How to connect AI agents to your POS, schedule, reviews, books, and website: start read-only, give each agent its own login, and approve anything public on the exact text.
To connect AI agents to your business tools, give each agent its own login through the vendor's connector or an MCP server, start it read-only, and add write access one action at a time. Anything a customer will see, like a website change, a Google post, or a review reply, should wait for a person's yes on the exact text. Keep a log of every call so you can see what the agent read, what it changed, and who approved it.
[VIDEO: 60 to 90 second screen recording. An agent drafts a reply to a new Google review, the owner reads the exact text, approves it, and the entry shows up in the activity log. Planned, not yet recorded.]
It's Tuesday, 10 PM, and you want one answer: how did labor run against sales tonight? Sales live in the POS, hours in the scheduling app, last week's comparison in a manager's spreadsheet. An agent can pull that together in seconds once it can reach those systems. Then you've got a new question. What else can it do in there?
I run a small fleet of AI agents across my own work, and getting them connected has never been the hard part. Deciding what each one is allowed to touch is.
What "connecting an agent" actually means
On its own, an AI agent can only talk. To read sales or draft a review reply, it needs a connection to the system holding the data. There are three common ways:
- The vendor's own connector. Built and kept up by the software company. Least setup; you get what they chose to expose.
- The vendor's API. Direct access, usually with a developer or integration platform in the middle.
- An MCP server. The Model Context Protocol is an open standard for connecting AI apps like Claude or ChatGPT to outside systems. Its own docs compare it to a USB-C port for AI (Model Context Protocol). A tool publishes one server, and any assistant that speaks MCP can use it.
For a restaurant, the vendor matters more than the method. Square publishes an official MCP server, still in beta, that signs you in with your Square account and lets you grant only the scopes you need (Square developer docs). Toast doesn't publish one. Its Standard API access, which an owner can turn on in Toast Web, hands out read-only credentials, and Toast says write access isn't included (Toast developer guide). So on Toast, an agent you connect yourself starts as a reader. That's a fine place to start.
Most independents run a stack of systems that rarely talk to each other (more on that in your restaurant runs on 10+ software platforms). An agent doesn't fix that by itself. It's one more thing with a login, so give it the right one.
What the agent should read, draft, and never send alone
| System | What the agent reads | What it can draft | Who signs off before it goes out |
|---|---|---|---|
| POS | Sales, voids, comps, item mix | The nightly summary | Nobody for an internal report. Nothing gets written back to the POS. |
| Scheduling | Shifts, hours, time-off requests | A proposed schedule change | The manager who owns the schedule |
| Reviews and Google listing | New reviews, rating, unanswered messages | A review reply or a Google post | You, on the exact text |
| Books and accounting | Invoices, vendor bills | Categories, a question for your bookkeeper | Your bookkeeper |
| Website and online menu | Pages, menu items, hours | An edit, a new page, an 86 | You, on the exact text |
Reading is cheap. Drafting is fine. Publishing is yours.
How to connect AI agents to your business tools in 7 steps
1. Start with one question you ask every week
Skip "connect everything." Pick a question you already ask a manager: last night's labor percentage, which reviews still need a reply, what's unpaid. That tells you which one or two systems to connect first. If it's the nightly numbers, the nightly sales report is a good spec for what the agent should hand you.
2. Use the vendor's connector before anything custom
If the software company ships a connector or MCP server, use it. They keep it working when their API changes, and their sign-in screen shows what you're granting. A community-built server means trusting whoever wrote it with your credentials.
3. Give the agent its own login, read-only
Don't paste the owner's password into an AI tool. Create a separate credential for the agent, limited to the locations and data it needs, with every write permission off for the first few weeks. If something looks wrong, you shut off that one login and nobody else gets locked out.
4. Don't trust a "read-only" label on its own
MCP lets a server label a tool read-only, but that's a hint. The MCP team's blog puts it bluntly: "A server can claim readOnlyHint: true and delete your files anyway" (MCP blog). One developer tested 19 maintained MCP servers in November 2026 and found that in two of them, a tool labeled read-only could write (DEV Community). The limit that holds is the credential's permission. If the login can't write, the agent can't.
5. Split writes into private and public
Once the reads are solid, add writes one at a time, sorted by who sees the result. A draft schedule a manager reviews is a private write. A website change, Google post, review reply, or menu price is public: the second it goes out, a guest reads it. Public writes stop at a draft and wait for your yes on the exact text. The MCP spec says the same thing in its own words: there should always be a human in the loop with the ability to deny tool calls (MCP specification).
6. Make the approval a record, not a message
"The agent said you approved it" isn't an approval. The system should record it when you press the button, tied to the exact change you saw. I learned this the hard way. Three times one afternoon, an AI reviewer approved a change in my setup, and three times my system refused to accept it, because the review record said it couldn't read a verdict. My first reaction was the one most people have: the tooling's broken, push it through. I couldn't, because I'd built it so I couldn't. It cost me an afternoon. It's also why I trust the approvals that do go through.
7. Keep a log and a budget, and read both every week
Every call should leave a line: which agent, which tool, what it read or changed, who approved it. Then cap each agent's spending. The agents that handle search work for the venues I work with are each capped at ten dollars a month per site. If one of them gets stuck in a loop, you get a small bill and a log entry to read, and nothing worse.
Before you connect anything: a checklist
- I can name the one question this agent answers, and which system owns each field it needs.
- The agent has its own login, not mine.
- That login starts read-only and only sees the locations it needs.
- I've tested in the vendor's sandbox where there is one. (Square recommends it. Toast's Standard API access doesn't include one.)
- Anything public, like the website, menu, Google listing, review replies, or social posts, stops at a draft.
- Approvals are recorded by the system when I press the button.
- Every call is logged, and each agent has a monthly spend cap.
- I can revoke the agent's login in under a minute.
What it looks like day to day
Most of what my agents produce is drafts. I approve what needs approving from my phone, often on the walk to the train. Approvals still need my yes on the exact text, and payments still need a tap only I can make.
That pattern holds whether the agent lives in Claude, ChatGPT, or Grok. The businesses on Apex Insights reach it through its MCP server from those assistants more often than through the app, and the approval step is the same from each. If you're also deciding whether to put one system beside your POS or switch registers, start with all-in-one restaurant software that isn't another POS.
An agent that can read your systems isn't the risk. An agent nobody decided the limits for is. Decide what it can write first, and connecting it is the easy afternoon.
Questions owners ask about connecting AI agents
Can an AI agent connect to Toast POS?
Not through an official Toast MCP server, because Toast doesn't publish one. Owners on Restaurant Management Suite Essentials or higher can create Standard API access credentials in Toast Web. Those are read-only, so an agent using them can read sales and labor but can't change the POS.
Does Square have an MCP server?
Yes. Square publishes an official MCP server, currently in beta. You sign in with your Square account and grant only the scopes you need. Square recommends testing your prompts against a Sandbox account before you point it at real data.
What is MCP in simple terms?
The Model Context Protocol is an open standard that lets AI apps like Claude or ChatGPT connect to outside systems. A business tool publishes an MCP server once, and any assistant that supports MCP can use its tools, within the permissions you grant.
Should an AI agent post review replies or website changes on its own?
No. Let it draft the reply or the change, then approve the exact text yourself. Anything a guest can read is public, and a wrong word there costs more than the minute it takes you to check.
How do I see what an AI agent did in my systems?
Read the activity or audit log in the agent tool and in each connected system. Each entry should show the agent, the tool it called, what it read or changed, and who approved it. If a connector has no log, keep it read-only.
Want help connecting your POS, schedule, reviews, and website to an agent, with nothing going out until you approve it? Get in touch with the Apex Insights team.