MCP explained: how we connect Claude to your systems.
The Model Context Protocol in plain words for a business owner: what an MCP connection is, what we connect in practice, what it changes for a team, how access is controlled, and how to start. Published 5 November 2026.
The short answer
MCP, the Model Context Protocol, is an open standard that lets Claude or ChatGPT use a system you already run, through a defined set of actions with permissions, instead of you copying and pasting between the two. An MCP connection is one such link to one system: your mailbox, Xero, the CRM, your own database. Sentry AI builds these connections as part of moving a team onto Claude, so the skills people use read and write real data, with access set per person and per department and anything that changes a record held for a named person to approve.
What an MCP connection is
Before MCP, getting an AI assistant to use your accounting system meant a custom integration, written once, for one assistant, that broke when either side changed. The Model Context Protocol replaces that with a standard. A small piece of software called an MCP server sits in front of one of your systems and describes, in a fixed format, what an assistant is allowed to do there: search these records, read this document, draft that reply. Any assistant that speaks the protocol can use it.
For a business owner the useful picture is a power socket. The system is the appliance, the connection is the socket, and Claude is whatever you plug in this year. The protocol is Anthropic's and it is open, so the sockets you build are not tied to one vendor's model. That matters more than it sounds: the connections and the skills that use them are where a business's AI value sits, and they should outlive any one model.
What we connect in practice
The systems below are the ones the teams we look after actually run. One client team has 12 data connections behind 45 skills; most start with two or three.
Microsoft 365 and Outlook
Mail, calendar, SharePoint and Teams. The connection a practice manager or an operations lead uses first, because that is where the day already lives.
Google Workspace
Gmail, Calendar, Drive and Slides, for the businesses that run on Google. The same shape of connection, the same permissions.
Xero
Invoices, contacts and the ledger, read in place. A skill can prepare a debtor follow-up or a month-end summary from the real numbers rather than from a pasted export.
Your ERP
NetSuite, MYOB Advanced, Dynamics or whichever runs the ledger and the orders. Management and finance reports read the real general ledger, so a month-end pack or a cash view is prepared from the numbers in place, with a person checking it before it goes anywhere.
The CRM
Whichever one the business runs. Deals, contacts and the notes behind them, so a skill that drafts a follow-up knows what was last said.
Shopify, Slack, Jira and the rest
Orders and products, the channels where decisions are made, the tickets where work is tracked. Each is one connection with its own permissions.
Your own database
The system that makes the business different, often older and without a connection anyone sells. We build the connection in your own repository, so you own it and it outlives us.
What it changes for a team
Without a connection, an assistant is a very good blank page. With one, the skills a team uses every day stop asking people to fetch things.
Skills read real data
A skill is the written instruction Claude follows for one job. With a connection, the weekly report reads the real ledger, the supplier email reads the real order history, and nobody pastes anything.
Skills can write, under control
Draft the reply in the mailbox, prepare the invoice in Xero, log the note in the CRM. The write is the step that saves the most time, and the step most carefully gated.
The same answer for everyone
When the connection and the skill are shared across the organisation, every seat gets the same result from the same source, and a fix to the skill reaches everyone the next time they use it.
Context that is yours
The connections, the skills and the pages they read are plain, portable and owned by the business. Change the model and nothing else moves.
The pages those skills read, how the business works written down once, are the second brain for your business. A connection is how the brain reaches the systems; a skill is what it does there.
Who may do what
How access is controlled
This is the question every owner asks second, and it has a plain answer at four levels.
Per person
A connection is added by the business's own administrator with the narrowest permission that does the job, and each person sees only the connections they were given. Finance connections are not open to the shop floor.
Per department
Skills and the connections they use are grouped by role, so reception, finance and the leadership team each get what their jobs need and nothing more. This is access control in the plainest form: who may use which skill against which system.
Approvals
Reading, analysing and preparing is the default. Anything that changes a system of record can be held as an approval that a named person accepts or rejects before it happens.
Nothing crosses the line
Every record belongs to one organisation. Connections, tokens and data never cross between businesses, and disconnecting a system removes its access at once.
Who is accountable for what an agent did, and how leadership keeps a register of every one and a way to stop it, is the subject of AI governance.
What MCP is not
Not a plug-in marketplace
You are not shopping for apps. A connection is built or configured for your business, against your systems, with permissions your administrator sets.
Not a product we sell
The connections are part of moving a team onto Claude. They live in your own accounts and repositories, and the value sits in the skills and knowledge around them, not in the plumbing.
About systems, not people
A connection carries the data a skill needs to do a job. What a person typed, and what the AI answered, is never captured by us.
How to start
If a connection already exists, start with a review of it: what it may read and write, who can use it, what it costs in usage, and whether the reports built on it are sound. If none exists, start with the Claude training day. Your team names the jobs, we build the first skills in the room, and the first connection is usually the system those jobs live in: the mailbox, the accounting system or the CRM. Nothing is connected until the person who owns that system has agreed what the connection may do, and the baseline for each job is timed before the connection exists, so the time it saves can be measured rather than claimed. The wider rollout is described on Claude for business.
Where the model itself has to run on your own hardware, the connections are the same; only the model behind them changes. That is the point of building them this way, and it is why our open weight models work and this page sit together.
Frequently asked questions
What is MCP, in one sentence?
The Model Context Protocol is an open standard, published by Anthropic, that lets an AI assistant such as Claude use a system you already run, through a defined set of actions with permissions, instead of you copying and pasting between the two.
What is an MCP server?
The small piece of software that sits between the assistant and one of your systems. It describes what the assistant is allowed to do there, such as read an invoice, search a mailbox or draft a record, and does only that. One connection per system; the assistant can be given several.
Which systems can Claude connect to?
In practice for the teams we look after: Microsoft 365 and Outlook, Google Workspace, Xero, an ERP such as NetSuite, Shopify, Slack, Jira, the CRM, and a company's own database. Where a connection does not exist for a system, we build one in the business's own repository, so the business owns it.
Is this safe? Can the AI change our records?
Only what the connection allows, and only for the person using it. Reading and preparing is the default. Anything that changes a system of record, from sending an email to updating a customer, can be held as an approval that a named person accepts or rejects. Each person sees only the connections they were given.
Does the business have to be on Claude?
No. MCP is an open standard and ChatGPT and other assistants read the same connections. Most of the teams we move over are on Claude, which is where the protocol started and where organisation-wide connections and skills are most mature, but the connection is the asset and it outlives the choice of model.
How do we start?
If a connection already exists, with a review of it: what it may read and write, who can use it, what it costs in usage, and whether the reports built on it are sound. If none exists, with the Claude training day. Your team names the jobs, we build the first skills in the room, and the first connection is usually the system those jobs live in: the mailbox, the accounting system, the ERP or the CRM. Nothing is connected before the person who owns that system has agreed what the connection may do.
We already have an MCP connection to our ERP. What should we check?
Four things: what the connection may read and write, who can use it, what each report actually reads (entity, period, posted or unposted), and what the usage costs per seat. That is what our review covers.
Can our own team change a connection or a skill?
A skill, yes: it is plain words and your people own it. The connection's permissions sit with your administrator. New connections are code, and we build them in your repository so you can take them on.
Which system would you connect first?
Send us your website and we build a Claude assistant for your business before the first call, so you see what a connection would change on your own work.
Ask an AI about this page
Opens the assistant with this page loaded: read it, summarise it, cite it.