MCP in 90 seconds
Model Context Protocol explained without the spec jargon. The 'why' before the 'what'.
MCP in 90 seconds
Imagine your AI agent is a smart but newly-hired analyst. They’re brilliant — but they don’t know your data, can’t open your apps, and have no permissions in your systems.
MCP is the standard hiring contract that gives that analyst access to specific tools and data — without giving them the keys to everything.
The 30-second version
MCP (Model Context Protocol) is an open spec, published by Anthropic in November 2024, that lets AI agents talk to tools and data sources in a standard way.
Before MCP, every agent integration was bespoke. With MCP, you write an integration once and any MCP-aware agent can use it.
It’s like USB for agents. Plug in any tool, any agent picks it up.
Why it matters
- Reusable integrations. Write an MCP server for GitHub once → Claude, Cursor, Continue, and every other MCP-aware host can use it.
- A better place to enforce access. MCP standardises the connection and tool-discovery surface. The actual permission model still depends on the host, server, and how you configure credentials — MCP doesn’t magically make unsafe tools safe, but it does give you a more consistent place to apply scopes.
- Cross-vendor support. Microsoft, Anthropic, and OpenAI all support MCP today. That’s unusual cross-vendor agreement, and it makes MCP the most likely-to-stick tool-connection protocol of 2026.
The 90-second mental model
There are three moving parts:
- MCP Host — the app the user interacts with. Claude Desktop, Cursor, Continue, etc.
- MCP Client — a component inside the host that maintains a connection to a server.
- MCP Server — the process or service exposing tools, resources, or prompts.
The protocol is the contract those clients and servers use to talk.
When you ask Claude to “summarise the open issues in repo X”, here’s what happens:
- Claude (the host) has the GitHub MCP server connected via its MCP client.
- Claude calls a function on the server:
list_issues(repo: 'X'). - The server returns the issues.
- Claude reads them and summarises.
You don’t write any of that wiring code. The protocol handles it.
The catch
MCP is still moving fast. Auth flows, transport (including streaming), and lifecycle behavior are evolving across spec revisions. Future updates may require migration work.
If you’re building today, prefer MCP servers from official vendors (GitHub, Cloudflare, Anthropic, etc.) over random hobby implementations. They track spec changes faster.
What to do next
- Read the Standards overview for an opinionated take on MCP and its peers.
- Browse the MCP Health Board — the servers worth your time.
- Try Claude Desktop — easiest way to see MCP working in practice.