What is MCP?
Updated November 2026
MCP, or Model Context Protocol, is an open standard for connecting AI applications to external systems. It provides a common integration boundary for capabilities such as tools, data resources, and reusable prompts rather than defining how multiple agents coordinate.
How does an MCP connection work?
The official MCP introduction describes connections to data sources, tools, and workflows. A host application uses an MCP client to communicate with a server exposing supported capabilities. The host still decides which capabilities to present to its model and how their results enter the conversation.
For example, a support assistant might read customer records through a server instead of embedding database-specific integration code throughout its reasoning workflow. Keep the exposed operation narrow: a lookup by customer identifier is easier to constrain than unrestricted query execution.
What belongs in MCP rather than A2A?
The A2A documentation distinguishes agent-to-tool communication from agent-to-agent communication. Use MCP when the useful contract is an operation or data interface. Consider A2A when the remote party is an independently operating agent responsible for completing delegated work.
| Interface | Responsibility |
|---|---|
| MCP tool | Expose a callable operation |
| MCP resource | Expose contextual information |
| Agent delegation | Assign work to another agent application |
Which controls does MCP not replace?
A standard interface does not make an operation safe. Authenticate the caller, scope credentials, validate arguments, and require approval for consequential writes. Treat tool descriptions and returned content as untrusted input, especially when they contain text that could influence later model instructions.
The tradeoff is integration consistency versus a broader capability surface. Connecting more servers increases available functions but also increases permission review, tool selection ambiguity, and exposure to malicious content. Start with a small allowed set and log actual calls, not merely the model’s stated intent.
What else should you know?
Is MCP an agent framework?
No. It is an integration protocol, not a complete agent runtime.
Does a server grant unlimited access?
It should not. Access depends on server controls and application authorization.
Can MCP support read-only use?
Yes. An application can expose or permit only read operations.
Does MCP approve actions?
No. Approval policy belongs to the application and its execution controls.