The concept
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems: data, tools and workflows.
The official site calls it "a USB-C port for AI applications". One standard port, and any charger fits any laptop.
The parts, in plain words:
- MCP server: a small program that wraps one system (say, a ticket tracker) and offers its abilities in the standard format.
- MCP host: the AI app you use, like a chat assistant or code editor.
- MCP client: the connector inside the host, one per server.
A server mainly offers:
- Tools: actions to call, like "create ticket" (Day 37).
- Resources: data to read, like a document.
- Prompts: reusable templates, like "summarize this sprint".
Walkthrough: a product owner's assistant
Maya, a product owner, wants her assistant to use the ticket tracker, the shared docs drive and the team chat.
Without MCP:
- Her company uses 3 AI apps (a chat assistant, a code editor, an internal bot).
- Each needs a custom integration per system: 3 × 3 = 9 connectors.
- Add a 4th system: build 3 more.
With MCP:
- One MCP server per system: 3 servers.
- Any MCP-compatible app can connect to all of them.
- Add a 4th system: build one server, every app gets it.
Maya asks: "Summarize open bugs for the payments epic and post it to the team channel."
- The assistant asks the ticket server which tools it has, then searches open bugs.
- It reads the epic's spec as a resource from the docs server.
- It drafts the summary.
- It calls the chat server's "post message" tool, after Maya approves.
The model still never touches the systems directly. The host and servers do the work.
A word of care
MCP standardizes the plug, not the trust. A server can do whatever its access allows. Use servers from sources you trust, apply least privilege (Day 38), and require approval before posting or deleting.
Where you see it at work
- Product management: one assistant reading tickets, specs and chat without copy-paste.
- Banking: a read-only server over the policy library, reviewed once, reused by several AI apps.
- Healthcare: a tightly scoped server exposing only approved, non-sensitive data.
Official site: https://modelcontextprotocol.io
In one sentence
MCP is a shared standard that lets any compatible AI app plug into tools and data through servers, so you build each connection once, not once per app.