If you have read anything about AI agents this year, you have run into three letters: MCP. The Model Context Protocol has quietly become one of the most important standards in AI — the "USB-C port" that lets any AI assistant plug into your tools and data. Here is what it is, why it matters, and how to use it.
The problem MCP solves
An AI model on its own only knows what it was trained on and what you paste into the chat. To be useful at work it needs to read your data (documents, tickets, databases) and take actions (create an invoice, update a CRM record, deploy code).
Before MCP, every AI product needed its own custom integration with every tool: one connector for ChatGPT and Slack, another for Claude and Slack, another for Copilot and Slack, and so on. With dozens of AI products and thousands of business tools, that does not scale.
What MCP is
The Model Context Protocol is an open standard that defines one common way for AI applications to connect to external tools and data. It was introduced by Anthropic in late 2024 and is now governed by the Agentic AI Foundation, a Linux Foundation project co-founded by Anthropic, Block and OpenAI (Anthropic).
The idea is simple: build an MCP server for a tool once, and any MCP-compatible client — ChatGPT, Claude, Gemini, Microsoft Copilot, Cursor, VS Code and many others — can use it.
How it works
MCP has three parts:
- Host / client — the AI application the user works in (for example Claude, ChatGPT or an IDE).
- Server — a small service that exposes a tool or data source in the MCP format.
- The protocol — the shared language between them.
An MCP server typically offers:
- Tools — actions the AI can call, such as "search orders" or "create ticket", each with a clear description and input schema.
- Resources — data the AI can read, such as files, records or documents.
- Prompts — reusable templates for common tasks.
When a user asks "What is the status of order 1042?", the AI sees that a connected server offers a "get order" tool, calls it with the order number, reads the result and answers — all through the same standard interface it would use for any other system.
How widely is MCP used?
By the time Anthropic donated MCP to the Linux Foundation in December 2025, there were more than 10,000 active public MCP servers and over 97 million monthly SDK downloads, and it had been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot and VS Code (Anthropic). In AI job postings, agent and MCP-related skills are among the fastest growing of 2025–2026 (Stanford AI Index 2026).
What changed in 2026
MCP is "growing up", as the Agentic AI Foundation put it. The 2026 specification work focused on making MCP easier to run at scale — moving towards a stateless design that fits ordinary web infrastructure, adding formal extensions such as MCP Apps (interactive UI inside AI clients) and Tasks (long-running work), retiring rarely used features, and tightening OAuth-based authorisation (AAIF).
MCP vs a regular API
| Regular API | MCP server | |
|---|---|---|
| Built for | Developers writing code | AI models deciding what to call |
| Descriptions | Docs for humans | Machine-readable tool descriptions |
| Reuse across AI apps | Custom integration each time | Works with any MCP client |
| Typical use | App-to-app integration | Giving assistants and agents tools |
MCP does not replace your APIs — an MCP server usually wraps an existing API so AI can use it safely.
MCP vs A2A
You will also hear about A2A (Agent2Agent). MCP connects an agent to tools and data; A2A lets agents talk to other agents, including ones from other vendors. A2A passed 150 supporting organisations in its first year (Linux Foundation). The two are complementary. Our guide to AI agents for business explains how they fit together.
How businesses use MCP
- Internal knowledge — let staff ask an assistant questions across your wiki, drive and ticketing system.
- Customer support — give a support agent read access to orders and shipping, and controlled write access to refunds.
- Sales and CRM — let assistants look up accounts, log calls and draft follow-ups.
- Engineering — connect coding agents to issue trackers, logs, databases and deployment tools.
- Your own product — publish an MCP server so customers can use your service from inside ChatGPT, Claude or Copilot. For SaaS companies this is becoming a new distribution channel.
Security: what to get right
MCP makes it easy to give AI powerful access — which is exactly why it needs care:
- Least privilege. Expose only the tools a use case needs; make read-only the default.
- Authenticate properly. Use OAuth and per-user permissions rather than shared admin keys.
- Trust carefully. Only install MCP servers from sources you trust, and review what each tool can do.
- Defend against prompt injection. Data returned by a tool (an email, a web page) may contain instructions; it must never be allowed to expand the AI's permissions. See our guide to AI security and prompt injection.
- Log every call so actions are auditable.
- Require confirmation for destructive or financial actions.
Getting started
- List the systems your team most often copies information in and out of.
- Check for existing MCP servers — many popular tools already have official ones.
- Build a server for your own system where none exists; it usually wraps an API you already have.
- Pilot with one team, with read-only tools first, and add write actions with approvals once it is trusted.
Zinetics builds MCP servers, AI agents and AI integrations for products and internal systems. If you want your platform to work inside ChatGPT, Claude or Copilot — or want an assistant that can actually use your tools — get in touch.