MCP is an open protocol that gives AI assistants access to data in business systems - without a manual integration for every tool. We explain how it works and why most business software will run on it within a few years.
Ask Claude or ChatGPT today about the status of a specific job, and the model will tell you it doesn't know - it has no access to your data. Not because it can't, but because there's no standardized way for it to ask your system for data. That's exactly the gap Model Context Protocol, or MCP, closes.
What MCP actually is
MCP is an open protocol Anthropic (the creator of Claude) introduced in November 2024 as a standardized way to connect AI assistants to the systems where data actually lives - CRM, databases, email, business software. Before MCP, every tool had to build its own hand-written integration for each AI application separately. Anthropic compares the protocol to a USB-C port - one connector instead of dozens of proprietary cables.
In under a year and a half, OpenAI, Google and Microsoft adopted it as clients too, so today it isn't an "Anthropic standard" - it's a shared language across the major AI platforms. In spring 2026, governance of the protocol also passed to the Agentic AI Foundation under the Linux Foundation - the same umbrella that hosts Kubernetes, for instance - a signal to businesses that this isn't one company's experiment, but a standard that will outlive a change of vendor.
How MCP works
A system that wants to expose its data runs an MCP server. In it, it describes three things: tools - actions AI can call, like "find a job by number" or "create a new contact"; resources - data it can read, like documents or records; and prompt templates - ready-made ways for AI to ask about a given thing. An AI assistant - the MCP client - connects to the server and uses these capabilities right away, without a developer having to write connection code tailored to a specific model.
Apexloop already uses this exact logic inside the app for the AI assistant over your workspace: reading data is unrestricted, but AI gives you every change to approve first. MCP carries this same principle beyond the boundaries of a single application - any AI agent can ask about data the same way, without automatically getting unrestricted write access.
How it differs from an API and a webhook
If you know the terms API and webhook, MCP will feel familiar - it builds on the same foundations. The difference is who it's built for. An API and a webhook connect two specific systems according to a fixed contract a developer designed on both sides. MCP is built directly for the AI model: the server describes its capabilities so the model can understand them on its own, without hand-written connection code for every pair of tools.
In practice, this means a company exposes its data through an MCP server once, and any AI assistant that supports the protocol can then connect to it - today's assistants and the ones that won't exist until next year.
Where you actually see it today
In under two years since its introduction, MCP has over 10,000 publicly running servers and more than 97 million SDK downloads a month. The latest version of the specification, from July 28, 2026 - the biggest revision in the protocol's history, according to its maintainers - also removed the need to maintain a stateful connection between client and server. That means MCP servers can run on ordinary, scalable web infrastructure instead of a specialized setup, sharply cutting the cost of running them.
Someone in Czechia is already doing this
Rohlík.cz, a Czech grocery delivery service, runs a public MCP server that lets a customer place an order right inside a conversation with Claude or another AI assistant - without opening the website or the app. It's still the exception, not the rule, but it shows the direction business software is heading overall.
What this means for a smaller company
MCP servers today are mostly built by large companies with their own development teams - connecting to the protocol means deciding which data and actions are worth exposing, and coding them safely. For a company using an off-the-shelf system instead of building its own, it mainly means one thing: once the vendor of its CRM, project management or business system decides to support MCP, its data suddenly becomes available to AI tools that don't even exist yet today - without the user having to build anything themselves.
Apexloop has already gone down this road: a single data model underlying CRM, jobs, projects and e-signatures turned out to be an ideal starting point for its own MCP server, which is running today - with the same write-approval rule that applies inside the app. How an agent actually connects to the workspace, and what it can read and propose, is shown in the article Apexloop as a data source for AI agents.
Access, not blind trust
MCP by itself doesn't decide whether an AI agent sees all the data or only part of it, or whether it can change records on its own without approval. That's down to how you configure the specific MCP server - just like with access rights inside the system. The standard gives you a shared language, not an automatically secure configuration.
How Apexloop handles who on a team sees which data today is covered in the guide Access rights and data visibility. Where your company data actually goes once you put it into an AI tool is covered in the article You upload company data into AI. Where does it actually go? You'll find a short definition of the term in the glossary too, and the news post covers the launch of our own MCP server.
Frequently asked questions about MCP and business data
Is MCP the same thing as an API?
Not quite. MCP often builds on an API, but it also describes its capabilities in a way an AI model understands directly - without hand-written connection code for every combination of tools. You'll find the difference between the terms in the glossary: what is an API, what is a webhook.
Does a small company need MCP too?
Building your own MCP server today mainly makes sense for companies with a development team. A smaller company benefits more indirectly - once the tools it already uses start supporting it, its data becomes available to AI assistants without any further effort on its part.
Is it safe to let an AI agent near business data over MCP?
It depends on how the specific server is configured - what data it exposes, whether it distinguishes reading from writing, and whether writing requires approval. The protocol itself doesn't guarantee security; that has to come from the operator's configuration.
Does Apexloop have an MCP server?
Yes. The Apexloop MCP server runs today, and an agent can use it to read workspace data and propose changes that wait for approval - exactly as described in the article Apexloop as a data source for AI agents.