MCP (Model Context Protocol)
Where it sits in the kitchen
The standard plug: any appliance built with it works in any kitchen that has the socket.
Definition
MCP, the Model Context Protocol, is an open standard for connecting AI apps to other software.
Anthropic, the company behind Claude, introduced MCP in November 2024. Since December 2025 it has been governed by the Agentic AI Foundation, part of the non-profit Linux Foundation. A company builds one MCP server, a program that offers its tools, data and ready-made prompts. Every app that supports MCP can then use it. Many apps call such servers connectors. Each server can reach the accounts you connect it to, so you must trust whoever made it.
Detailed technical description
MCP defines how a host (the app running the agent, see harness) connects, through MCP clients (the part of the host that talks to one server), to MCP servers. Servers offer tools (actions), resources (data) and prompts (templates). They run on your own computer or remotely over the web (HTTP), with access granted through OAuth, the standard behind “allow this app access” screens.
Since its launch, MCP, an open standard (a public format any app can use), has been adopted across the industry. The current revision (2026-07-28) simplified the core so that a server no longer has to remember each client between requests, which lets remote servers spread across many machines. It moved optional features, such as long-running tasks and interactive interfaces (MCP Apps), into formal extensions.
Use MCP for integrations many agents can reuse (GitHub, Jira, Slack, databases, internal systems) and to make your own product usable by AI agents. Do not load many servers with hundreds of tools into every chat or job: use tool search, or code execution with MCP, where tools appear as code so large intermediate results never enter the context. Treat every server as untrusted code that holds your credentials (logins and keys): check who publishes it and what access it asks for, stay on a version you have checked, review changes to its tool descriptions, and use OAuth 2.1 with per-user, least-privilege scopes, so each user grants only the access the task needs.