Somewhere in the release notes of your accounting package, between a bug fix and a new invoice template, a line has probably appeared this year: MCP connector now available. Your CRM may have announced one too, and your webshop platform almost certainly has. Nobody sent you a memo explaining what that acronym means — yet it is quietly becoming the plumbing that decides what AI can and cannot do inside your business.
The USB-C of AI, in one paragraph
MCP stands for Model Context Protocol, an open standard that Anthropic released in November 2024. Before it, every connection between an AI model and a business tool was a custom job: one integration for this assistant and that CRM, another for that assistant and this spreadsheet. MCP replaces the tangle with a single plug. A tool exposes an MCP server — a standard description of what it can show and do — and any MCP-capable AI can connect to it, the way any USB-C cable fits any USB-C port. Build the connector once, and it works with Claude, ChatGPT, Gemini and whatever comes next.
From side project to industry standard in twenty months
The adoption curve is the story. OpenAI — Anthropic's chief rival — adopted MCP in its Agents SDK in March 2025 and later joined the protocol's steering committee. Google committed Gemini to it in April 2025. In May 2025, Microsoft made it a first-class standard across GitHub, Copilot and Windows 11, and AWS built support into its Bedrock agent platform that summer. In December 2025, Anthropic handed the whole protocol to the newly created Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI — meaning no single vendor owns the plug anymore.
The ecosystem followed. An official MCP registry launched in September 2025 and, according to identity firm WorkOS, grew to roughly 2,000 listed servers within months, while the protocol's Python and TypeScript toolkits were being downloaded about 97 million times a month by spring 2026. Among the tools with official or community connectors: Shopify and WooCommerce for e-commerce, Stripe for payments, Salesforce and HubSpot for CRM, plus Slack, Notion, monday.com and Canva.
Why this month matters
On 28 July the protocol's biggest overhaul yet becomes final: the 2026-07-28 specification, published as a release candidate on 21 May and hardened through a ten-week validation window with the companies running MCP in production. The changes read technical — the protocol becomes stateless, gains a Tasks extension for long-running jobs, and gets stricter authorisation rules — but the translation is simple: MCP was redesigned so that big companies can run thousands of agent connections reliably and securely. When AWS, Cloudflare, Google Cloud and Microsoft publicly commit to a version, as they have to this one, software vendors build on it with confidence. That is why the connector announcements are accelerating, and why more will land in the tools you already pay for.
What a connected agent can actually do for you
The point of all this plumbing is mundane and useful. Without MCP, an AI assistant knows only what you paste into it. With connectors, an agent can look things up and act across your systems in one instruction.
Accounting. Ask which invoices are more than 30 days overdue, and get the list pulled live from your books — plus drafted payment reminders in your tone, ready for your approval.
E-commerce. Have the agent check which products ran out of stock this week, cross-reference the bestsellers, and prepare the reorder list — reading your actual shop data rather than guessing.
CRM. A Monday-morning summary of every open deal that has gone quiet for two weeks, with a suggested follow-up for each, drawn from the real pipeline.
The pattern: the agent stops being a well-read intern with no building access and becomes one with exactly the keys you chose to hand over.
The five questions to ask before you connect anything
Which brings us to the part that deserves slow thinking, because a connector is an access grant, not a feature toggle.
Who built it? An official server from your software vendor beats a community-built one, especially for financial data. The registry lists both; they are not equivalent.
Read or write? A connector that can only look things up is a far smaller risk than one that can send, edit or pay. Start read-only; add write access per task, not by default.
Where does the data go? Customer data flowing through a connector to an AI provider is a GDPR processing activity. Check that a data-processing agreement covers it and note it in your records of processing — same discipline as any other cloud tool.
Can you see what happened? Ask for logs of what the agent accessed and did. WorkOS's 2026 review notes that standardised audit trails are still a gap in the ecosystem — which makes vendors who provide them worth preferring.
Can you pull the plug? Know where connections are managed and who in your business can revoke one in two minutes.
When to wait
If your core tools have no official connector yet, resist the temptation to wire a random community server into live financial or customer data — with a maturing standard, six months of patience is often the cheapest security measure available. And if you have no recurring task that would use the connection weekly, you are adding attack surface without adding value. Connect deliberately or not yet.
MCP will keep spreading whether you engage with it or not; the July release exists precisely because large firms already run it at scale. The useful move for a small business is not to study protocols but to know which of your systems hold the data your future agent will need, and in what state. Mapping that — tools, data quality, permissions, and the tasks worth connecting first — is exactly what Cresly's AI Readiness Scan does, so that when your software announces its MCP connector, you already know whether, and how, to say yes.