MCP (Model Context Protocol) is the open standard that lets AI agents call external tools — databases, APIs, search engines, file systems — without each agent needing its own custom integration for each tool.

Before MCP, every AI tool had its own way of connecting to external services. MCP is the USB standard for AI tools: plug in once, works everywhere. An agent that speaks MCP can call any MCP server, and an MCP server can be called by any agent — so 5 agents and 10 tools becomes 15 things to build, not 50. It works through three components: a Host (the AI agent), a Client (the bridge inside the host), and a Server (the tool provider).

The problem MCP solves

Imagine you have 5 AI coding assistants and 10 external tools (GitHub, Jira, web search, databases…). Without a standard, each assistant needs its own custom integration for each tool: 50 custom integrations to build and maintain. With MCP, each of the 5 agents learns the protocol once, and each of the 10 tools exposes it once — 15 things to build instead of 50. The combinatorial explosion disappears.

How MCP works

MCP has three components:

  • MCP Host — the AI assistant or agent that wants to use tools (e.g., GitHub Copilot in agent mode)
  • MCP Client — the bridge inside the host that speaks the MCP protocol on the host’s behalf
  • MCP Server — the tool provider that exposes capabilities (e.g., a GitHub Issues server, a web search server)

The host asks the client; the client calls the server; the server responds with results. The agent never needs to know the implementation details of the tool — only the protocol.

MCP was originally developed by Anthropic and is now an open standard. Multiple AI tools — including GitHub Copilot — have adopted it.

What MCP servers can expose

MCP servers can surface three types of capability to agents:

  • Tools — callable functions the agent can invoke (search, create, update, delete)
  • Resources — data the agent can read (files, database records, documentation)
  • Prompts — reusable prompt templates the agent can invoke

In practice, tools are the most commonly used. When you add an MCP server to your environment, you’re primarily giving the agent new functions it can call.

⚠️ An MCP server is NOT an agent. It doesn’t decide anything. It doesn’t act on its own. It responds when called. The agent is the one with a goal. The MCP server is the one with a drill.

MCP in practice

GitHub Copilot supports MCP servers in agent mode. You configure them in your environment, and Copilot’s agent can call them during tasks. See Extensions, Tools & MCP for how to set this up in Copilot specifically.

Enterprise and Business users: MCP server support is disabled by default in Copilot Enterprise and Copilot Business. An administrator must explicitly enable the “MCP servers in Copilot” policy before users can connect MCP servers.

Dive deeper

Related Concepts

📍 Layer 4b — Tools/MCP · Back to the AICA Onion

What’s next?

  • Skills & Tools — how tools and skills fit into the bigger agent picture

Updated: