Model Context Protocol (MCP)
An open standard that lets models call tools and read resources exposed by any compatible server, across hosts.
Preparing spaced-repetition scheduler…
Before MCP, every agent framework had its own tool-calling convention. Model Context Protocol (MCP) is an open standard introduced by Anthropic in late 2024 that defines a single contract between a host (the model-bearing app) and one or more servers (tools, data sources, or context providers).
Three roles:
- MCP Host — the application that owns the model. Examples: Claude Desktop, an IDE plugin, a chat product.
- MCP Client — a connector inside the host. There is one client per server the host talks to.
- MCP Server — a process that exposes tools (callable functions), resources (read-only data), and prompts (reusable templates) over JSON-RPC.
Why it matters:
- Reusability. A server can be reused across compatible hosts. Transport support, protocol versions, authentication, and supported features must still match.
- Security boundary. Hosts enforce consent and access policy. A local server is executable software; using MCP alone does not sandbox it or make its results trustworthy.
- Composability. A host can mix local servers (filesystem) with remote servers (a hosted database) through the same protocol.
MCP messages use JSON-RPC. The standard transports are stdio for locally spawned servers and Streamable HTTP for remote servers; the older HTTP+SSE transport is retained only for backward compatibility.
Key takeaways
- MCP defines host ↔ client ↔ server over JSON-RPC.
- Servers expose tools, resources, and prompts.
- It is the open standard that decouples agent frameworks from tool integrations.
Self-check
What three primitives does an MCP server expose?
What transport does MCP use?
References
- Model Context Protocol — Introduction
Specification for connecting models to tools and data sources.
- Model Context Protocol — Transports
Current standard transports and HTTP+SSE compatibility guidance.
Last verified 2026-09-04.