Model Context Protocol (MCP)
MCP is a standard way for AI products to discover and use external data, software, and actions. This guide explains the protocol, the terminology used by major AI tools, and the practical work required to implement it safely.
Start here
How MCP works
Understand hosts, clients, servers, tools, resources, prompts, and the request lifecycle.
Stateless MCP
Learn how self-contained requests, explicit state handles, tasks, and subscriptions fit together.
Terminology
Translate between apps, connectors, plugins, integrations, and MCP servers.
MCP vs. APIs
See why MCP complements existing APIs instead of replacing them.
MCP in 30 seconds
MCP adds a shared contract between an AI host and the systems it needs to use. Your existing application still owns its data, permissions, and business logic.
-
1
Describe capabilities
A server exposes focused tools with clear inputs and outputs.
-
2
Select the relevant tool
The AI host chooses a capability while applying identity, permissions, and approval rules.
-
3
Return a structured result
The connected system performs the work and returns data the AI can explain, cite, or present for review.
Implementation guides
Implementation guide
Move from one useful workflow to a secure, observable MCP implementation.
Designing tools
Create tools that models can select correctly and people can understand and trust.
Progressive tool discovery
Keep large tool catalogs from crowding the model’s working context.
Security and control
Plan identity, least privilege, approvals, tenant separation, and auditability.
Product guides and examples
Need help implementing MCP?
Read about MCP development and hosted MCP, or review the sources and research policy behind this documentation.