MCP (Model Context Protocol) gives AI agents callable tools — search engines, databases, APIs. The agent decides when to call them. MCP is the tool layer, not the agent itself.

30-second version: When Copilot agents need to do things beyond editing code — searching the web, reading a database, creating a Jira ticket — they use tools. MCP is the open standard that defines how agents discover and call external tools. You configure MCP servers; the agent decides when and how to use them.

⚠️ An MCP server is NOT an agent. An MCP server is a tool provider — it exposes capabilities that an agent can call. The agent is still making the decisions. The MCP server just executes what the agent asks. Calling an MCP tool does not make the tool an agent.

What is MCP?

MCP (Model Context Protocol) is an open standard — originally developed by Anthropic — that defines a common way for AI agents to discover and call external tools. It works like a plugin system: you connect MCP servers to your Copilot environment, and the agents can call them during tasks. Think of it like a USB-C port for AI: the connector is standardised so any compatible tool can plug in.

GitHub Copilot supports MCP in agent mode. MCP support in enterprises is disabled by default and requires explicit policy enablement.

What can MCP servers do?

  • Search the web or internal documentation
  • Query databases or APIs
  • Read/write to project management tools (Jira, GitHub Issues, Trello)
  • Execute code or run tests in sandboxed environments
  • Access file systems or cloud storage

How Copilot uses MCP

When you use agent mode with MCP servers configured, the agent can autonomously decide to call a tool mid-task. You ask it to “fix this bug and create a GitHub issue for the root cause” — it edits the code AND calls your GitHub MCP server to create the issue. You configured the server; the agent decided when to use it.

GitHub maintains its own GitHub MCP server that exposes GitHub API operations as callable tools — creating issues, reading PRs, managing repositories — directly from inside agent mode.

Third-party agents (such as Anthropic Claude coding agent and OpenAI Codex agent) are accessible through GitHub Copilot’s orchestration layer and can use MCP servers the same way the first-party Copilot agent does. Third-party agents in GitHub Copilot are in public preview.

What happened to Copilot Extensions?

GitHub Copilot Extensions were deprecated in November 2025. MCP is the replacement. The official GitHub documentation for Copilot Extensions now redirects to MCP. If you see references to “extensions” in older documentation or blog posts, they map to what MCP servers do today — and MCP is the standard going forward.

Dive deeper

Related Concepts

  • What is MCP? — The conceptual foundation: what MCP is and the problem it solves
  • What is an Agent? — Agents are the hosts that call MCP tools to act on the world
  • Copilot Agent Mode — The Copilot context where MCP servers are actually used

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

What’s next?

  • Instruction Files — the other half of Layer 4: persistent context files that shape agent behaviour

Updated: