MCP (Model Context Protocol) — AI Glossary

Reference details

Related terms

• AI Agent
• Tool Use
• Function Calling

How it works

1. Client and server exchange JSON-RPC messages over a transport that is either a local process's standard streams or an HTTP connection.
2. The session opens with a capability negotiation, so each side learns what the other supports before any tool is listed and a newer client can talk to an older server.
3. The client then asks the server to list what it offers, and those returned descriptions and schemas are what the client injects into the model's tool definitions.
4. The model therefore never speaks the protocol: it emits an ordinary tool call, and the client translates that into a protocol invocation on its behalf.
5. Servers run as separate processes holding their own credentials, which is what makes a server a trust boundary — the client grants it access by starting it, not per call.
6. Because discovery happens at connection time rather than at build time, a server that changes what it exposes changes what the model can do without any client being rebuilt.

Example: N plus M instead of N times M

Four clients each needing five integrations is twenty bespoke connectors. With a shared protocol it is four clients and five servers — nine components, and a new server works with every client immediately. That arithmetic is the entire argument for standardising the interface.

bespoke : 4 clients x 5 tools = 20 integrations
protocol: 4 clients + 5 servers =  9 components

Commonly confused with

MCP is not function calling and does not replace it. Function calling is how a model expresses an intent to call a tool; MCP is how a client discovers and connects to the tools available. They operate at different layers and are normally used together.

When to use it

Reach for it when:

• The same tools must be reachable from several clients
• You want tool discovery rather than hard-coded integrations
• Tools are maintained separately from the applications using them

Reach for something else when:

• One client, two tools — direct integration is simpler
• You have not scoped what an MCP server exposes; it is a new trust boundary
• The tool surface changes rarely enough that discovery adds nothing

Referenced by

This term is used by the following ModelRefs references:

• mcp server integration

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to MCP (Model Context Protocol) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path