MCP explained for business owners: why your tools finally talk to AI
By Xavier PeichMCP, the standard that plugs AI agents into your software. What it changes for an SMB, and the questions to ask your software vendors.
For years, the most expensive part of a business AI project wasn't the intelligence. It was the plumbing. Getting an assistant to read your CRM, your invoicing system or your calendar meant a small custom IT project for each connection: built once, maintained forever. If you are shopping for a custom AI agent for your SMB, one acronym now comes up in every vendor conversation: MCP.
This article explains what it is, without a single line of technical detail, because what matters about MCP is economic rather than technical. It shifts the cost structure of AI projects, and it changes what you should demand from your software vendors, starting at your next renewal.
The short answer, for the busy
The Model Context Protocol (MCP) is an open standard that defines how an AI agent connects to external software: CRM, accounting, calendar, invoicing. Launched by Anthropic in November 2024, adopted by OpenAI in March 2025 and by Google in April 2025, it was handed in December 2025 to a neutral foundation under the Linux Foundation, which counted more than 10,000 public MCP servers at the time. Before MCP, every connection between an AI and a tool required a custom integration, built and then maintained by developers. With MCP, the software vendor publishes a standard socket, the MCP server, and any compatible agent can plug into it. For an SMB, the consequence is concrete: when your tools offer an MCP server, an AI agent can read them and act in them without a costly integration project. Stripe, Shopify, Notion and HubSpot already ship an official one.
Before MCP: plumbing you paid for twice
Take a simple problem: you want an AI assistant to answer customer emails by checking the history in your CRM and the invoice status in your accounting system. Before 2025, a vendor's honest answer sounded like this: the AI engine exists, but we have to hand-build a pipe to your CRM, then another to your accounting, then a third to your inbox.
Each pipe was paid for twice. Once to build it, then continuously: whenever the software vendor changed its system or the AI model evolved, the pipe broke and someone had to fix it. Multiply by the number of tools in an ordinary SMB and you understand why so many AI projects died at the quote stage.
That plumbing carried a second, less visible cost: dependence. The pipes belonged to the vendor who built them. Switching agent providers meant throwing away the integrations and starting over. Businesses bought custom for lack of an alternative, not out of any taste for artisanal plumbing.
The standard socket: what MCP fixed
MCP fixes this in the most mundane way possible: by standardizing the socket. Instead of building one pipe per software-AI pair, the software vendor publishes a single standardized interface, called an MCP server. Any compatible agent can plug into it, the way any appliance plugs into a wall outlet without negotiating with the utility company.
Anthropic published the standard in November 2024. What turned it into an industry norm is what came next: OpenAI, a direct competitor, adopted it in March 2025, and Google followed in April 2025 for its Gemini models. Then, in December 2025, Anthropic handed the protocol to the Agentic AI Foundation, a neutral body under the Linux Foundation, co-founded with OpenAI and Block. The announcement counted more than 10,000 active public MCP servers.
That last move deserves a pause, because it explains why you can build on this. Nobody invests in a competitor's proprietary socket: as long as MCP belonged to Anthropic, every software vendor hesitated. By ceding it to a neutral foundation, Anthropic traded control for adoption. That is the classic trajectory of infrastructure: email, the Web and the shipping container all won the day nobody owned them anymore.
What it changes for your SMB
The practical consequence fits in one sentence: your software becomes agent-accessible without an integration project, provided the vendor publishes an MCP server.
And the vendors SMBs actually use have moved. Verified in July 2026: Stripe offers an official hosted MCP server covering payments, invoices and subscriptions. Shopify publishes servers for a store's catalog, cart and customer accounts. Notion offers a hosted server covering the whole workspace. HubSpot ships one that reads and writes CRM data: contacts, deals, activities. The list grows every month.
Concretely, for the agent handling your customer emails: checking an invoice in Stripe or pulling up a file in HubSpot is no longer months of development, it is a connection. In our article on how an AI agent gets built, tools are one of an agent's six ingredients; MCP is turning that ingredient from custom work into a standard part. The remaining work shifts to what is genuinely yours: the mandate, the instructions, the guardrails.
Two economic effects follow. The entry cost drops, since the plumbing disappears from the quote. And dependence drops too: a standard connection belongs to no one, so switching agent providers no longer means rebuilding everything. Both effects work in your favour, which is rare enough to be worth noting.
The question to ask your software vendors
This shift moves a purchasing question. Ten years ago, you asked a software vendor: "do you have an API?". The 2026 question is: "do you offer an official MCP server?". Ask it at your next CRM, accounting or booking-system renewal. A vendor with no answer condemns you to custom plumbing, or to waiting.
But a "yes" needs probing, because official does not mean "plug and forget". The clearest example is QuickBooks: Intuit does publish an official MCP server, but it installs locally and requires developer configuration, far from the one-click connection of Stripe or Notion. Other servers cover only part of the product, or reading but not writing. Three follow-up questions sort it out: is the server hosted by the vendor or self-installed? Does it cover the data your agent will need? Does it allow reading only, or acting too?
This level of detail may sound technical. It is no more technical than asking whether a commercial space has power at the right voltage. You don't need to understand the protocol; you need to know whether the socket exists and what it feeds.
Security: what the agent touches, and on whose behalf
A standard socket makes connections easier. It does not decide what gets plugged in, and this is where the owner keeps a job nobody else will do.
Two questions frame everything else. First: what can the agent touch? Reading a customer's history is harmless; issuing a refund in Stripe or editing a CRM record is not. The rule we apply in our projects: an agent starts read-only and earns the right to act, permission by permission, never the reverse. Second: with whose permissions? An MCP server connects on behalf of an account you authorize. If that account sees everything, the agent sees everything. The right reflex is the one you apply to a new employee: their own access, limited to their mandate, revocable, with a log of what they did. And as soon as the data involved includes personal information, Quebec's Law 25 governs what you do with it: the connection standard changes nothing there.
None of these questions requires understanding the protocol. They require knowing what can actually hurt your business, and that is your job.
Where to start
The sensible sequence is three moves. Inventory your main tools and check which ones offer an official MCP server: the answer dictates what an agent can reach without extra cost. Ask the MCP question of any software vendor you are about to renew. Then pick one task, just one, where an agent plugged into your real tools would create measurable value: our primer on what an AI agent is for a Quebec SMB is the refresher if the concept needs one, and our comparison of no-code versus custom agents helps choose the vehicle.
That is exactly the inventory we do in a first conversation: your tools, their sockets, the task that justifies an agent, and an honest opinion on whether it is worth it. 30 minutes, no commitment.
Written by