What is Model Context Protocol (MCP)?
Model Context Protocol (MCP) is an open protocol, introduced by Anthropic in November 2024, that standardizes how AI applications connect to external tools and data. An MCP server exposes tools, resources and prompts. An MCP client inside an AI application calls them, so one integration works with any MCP-compatible app or model.
For example, a developer adds a CRM MCP server to the company’s AI assistant. A sales manager asks which renewal customers have open support issues. The assistant calls the server’s search tool, which returns 40 accounts with contact names, emails, phone numbers and contract values. All of that lands in the model’s context, even though the sales manager only needed the account names.
How MCP works
- Host. The AI application a person works in, such as a chat app, an IDE or an AI agent runtime.
- Client. The connector inside the host. Each client keeps a connection to one server.
- Server. A program that exposes capabilities from a system such as a CRM, database, file store or SaaS app.
- Tools. Functions the model can call, such as “search_accounts” or “create_ticket”.
- Resources. Data the application can read, such as files or records.
- Prompts. Reusable templates a server offers for common tasks.
Messages use JSON-RPC 2.0. Local servers usually run over standard input and output, and remote servers over HTTP. A simplified tool result might look like this:
{"content": [{"type": "text", "text": "Account: Northwind, contact: Ana Ruiz, ana.ruiz@northwind.com, +1 415 555 0142"}]}
Before MCP, every AI application needed its own connector for every system it used. MCP turns that into one server per system that any compatible application can reuse. That convenience is also why MCP servers spread fast inside companies, often with broad credentials and little review of what data they return.
MCP vs a regular API integration
| Regular API integration | MCP | |
|---|---|---|
| Built for | One app calling one service | Any MCP client calling any MCP server |
| Who picks the call | Code written by a developer | The model, from tools listed at run time |
| Where results go | The app’s own code | Usually straight into the model’s context |
Data risks in MCP
- Tool outputs often return whole records when the task needs one field.
- Tool arguments can carry sensitive values out to third-party servers.
- Many servers use one service credential for every user, so results ignore who is asking.
- Tool descriptions and returned content can carry prompt injection.
- Requests and results are often logged in full by the host, the server or both.
The post on handling sensitive data in MCP systems goes deeper, and data policy at the MCP gateway shows where to apply controls.
How Protecto helps
Protecto applies data policy to MCP traffic. It detects sensitive values in tool arguments and tool outputs, replaces them with consistent tokens before they enter the model’s context, and reveals real values only to the users and tools that policy allows. The guide on how to stop AI agents and MCP from leaking PII shows the pattern step by step.
Frequently asked questions
Who created the Model Context Protocol?
Anthropic introduced MCP as an open standard in November 2024. The specification and SDKs are open source, and many AI applications and model providers now support it.
What is the difference between an MCP client and an MCP server?
An MCP server exposes tools and data from a system such as a CRM or database. An MCP client lives inside the AI application and connects to a server so the model can use what it offers.
Is MCP secure?
MCP defines how AI applications and tools talk to each other. How safe a deployment is depends on which servers you connect, the credentials they use and the data they return, so treat every tool output as data that may reach the model and its logs.
Related terms
AI agent · Agentic AI · Prompt injection · LLM data leakage · AI gateway