Model Context Protocol (MCP)
- Agents
- Standard term
Model Context Protocol (MCP) is an open standard for connecting AI apps to programs that offer outside tools, data, or prompt templates. Those programs are called servers; compatible apps use the protocol's shared messages to communicate with them.
A comparison: Like USB-C's shared connector shape: a common connection reduces custom adapters, but a matching plug alone does not establish support for every feature.
An MCP server may run on your computer or remotely. An app can ask which capabilities it exposes; tools are described by names, descriptions, and accepted inputs. The model can propose a tool request, while the app and server govern whether and how it runs.
Where you’ll meet it
A message in your team chat announces an MCP server for the company's customer database, with steps for adding it to Claude or ChatGPT from the app's connector settings. When you add it, its list of tools includes searching customers, updating records, and deleting them.
What to do
For lookup work, ask whoever manages the server for read-only permissions in the database or service itself. Hiding write tools in an app can narrow what the model sees, but does not substitute for those permissions. Try a known search to check the useful access.
Easy to mix up
A connector is one app's saved link to a service, often running through an MCP server. MCP is the shared format that servers and AI apps use to talk to each other.
Check yourself
Someone shares an MCP server for your calendar. Its only calendar tool is named “t2” with the description “runs action,” and its listed inputs give no clearer purpose. You ask your assistant to add a dentist visit on Friday. Which problem remains?
Go deeperEffective access depends on server and client permissions
Anthropic introduced MCP as an open standard in November 2024, and other AI companies, including OpenAI, added support in 2025. Its messages use JSON-RPC, a simple format for requests and replies, and run either between programs on one computer or over the web.
A server's backend account sets an upper bound, while its own authorization rules and the client's configuration can restrict access further. An administrator account does not automatically give every client administrator powers. Tool descriptions are also information the model reads, so malicious wording can try to redirect it; this is known as “tool poisoning.”
Sources
- Introducing the Model Context Protocol, Anthropic
- New tools and features in the Responses API, OpenAI
- Model Context Protocol specification, Model Context Protocol
- MCP transports, Model Context Protocol
- MCP Security Best Practices, Model Context Protocol
- MCP Tools, Model Context Protocol
- MCP Security Notification: Tool Poisoning Attacks, invariantlabs.ai