Model Context Protocol (MCP)
MCP is an open JSON-RPC protocol that lets any AI host connect to tool and data servers through one interface. Now governed by the Agentic AI Foundation.
Evolving: the basics hold, details shift. Reviewed 17 Oct 2026.
Key ideas
- Roles: host (the app), client (one per server connection), server (exposes capabilities).
- Server primitives: tools (actions), resources (readable data), prompts (templates).
- Transports: stdio for local servers, Streamable HTTP for remote servers.
- Remote servers authenticate with OAuth 2.1; the server is a resource server and validates tokens.
- Design MCP tools with the same rules as any agent tool: task-shaped, well described, compact outputs.
- Test with the MCP Inspector before connecting to a host.
Common mistakes
- Passing user tokens straight through to downstream APIs (token passthrough is forbidden by the spec).
- Trusting tool descriptions from third-party servers; they can carry prompt injection.