MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to external systems. Through MCP, an application such as Claude or ChatGPT can connect to data sources like local files and databases, tools like search engines and calculators, and workflows like specialized prompts.
The USB-C comparison
The MCP project describes it as a USB-C port for AI applications. Just as USB-C gives a standard way to plug in different devices, MCP gives a standard way to plug an AI application into different systems. Build a connector once, and any client that speaks the protocol can use it.
The project's documentation lists examples: an assistant reading a calendar and notes, a coding agent working from a design file, and an enterprise chatbot querying several databases.
Who benefits
According to the project, developers spend less time on integration, AI applications gain access to an ecosystem of tools, and end users get more capable assistants that can act on their behalf when needed.
That last phrase is where security enters.
Why it needs a security review
An MCP server is code that sits between a model and your systems. Once an agent is connected, it can read data and take actions. That raises practical questions:
- Who wrote this server, and can you verify it?
- What can it reach, and does it need all of that?
- What happens if the model reads hostile text, as in prompt injection, and then calls a tool?
- Who approves actions that cannot be undone?
A short checklist before connecting a server
- Keep an inventory of every MCP server in use, including ones developers added themselves.
- Prefer servers from sources you can identify, and pin the version.
- Grant the narrowest permissions that work, with read-only where possible.
- Require human confirmation for destructive or outbound actions.
- Log tool calls so you can reconstruct what an agent did.
Where policy fits
Decide per server and per tool what is allowed, instead of trusting the whole connection. The project's own documentation includes a security section worth reading before rollout.
Click through it at your own pace, or open it full screen.
Safeguard offers an MCP Guardrail and agent identity to control which servers and tools agents may use. A free scan is at gold.safeguard.sh/scan.
Quick check
Is MCP a product from one company?
The project describes it as an open-source standard supported by many assistants and development tools.
Does MCP make an agent unsafe by itself?
No. It gives an agent reach. Safety depends on what you connect and what you permit.
Sources
Last reviewed: 19 October 2026.
