One standard plug that lets any AI use your tools. Like USB-C, for AI.
3 min to watch1 min to read
Based on real events: Anthropic released MCP in November 2024, and OpenAI adopted it in March 2025.1,2 The messages follow MCP's official spec.3 The screens, names and sales numbers are made up.
Read the walkthrough as text
Without a standard plug
A manager uses an AI assistant at work. It's good with words, but it lives inside this chat.
The manager asks for something that needs two other apps.
The AI says yes.
Behind the scenes, the assistant tries to reach the sales app. All it finds is a sign-in page.
Slack is closed to it too. So it has to tell the manager it's stuck.
The old fix: an engineer writes a custom connector, made only for this assistant and the sales app.
Then a second one for Slack, with its own code.
Then the company adds two more AI apps and two more tools. Every pair needs its own connector.
Three AI apps times four tools: 12 custom connectors to build, pay for, and fix.
One custom cable for every pair.Before MCP, each AI app needed its own connector to each tool, so the work multiplied. Anthropic released MCP in November 2024 as one open standard for these connections.
With a standard plug
Same request. This time the sales app and Slack each come with one standard plug for AI.
Behind the scenes, the sales app now has a plug made for AI apps, in a standard shape.
The assistant asks the plug what it can do. It gets a list of tools, in a format every AI app understands.
It picks that tool and calls it, with the details filled in.
The answer comes back as plain text the AI can read.
Slack's plug works the same way: list the tools, then call one.
This is MCP: one standard plug per tool, and any AI app can use it. Developers call each plug an MCP server.
Back in the chat, the job is done.
Build the plug once. Use it everywhere.Anthropic released MCP in November 2024, and OpenAI adopted it in March 2025. The tools/list and tools/call messages come from MCP's official spec.
Keep going:
Why it matters, what it costs, and what to ask your team.
Try it
Count the cables.
Without a standard, every AI app needs its own connector to every tool. With MCP, each one plugs in once. Add apps and tools:
3AI apps4tools
Custom cables
With MCP
Fewer cables means fewer things to break.
How it works
One request, step by step.
MCP is a shared language. The AI asks what tools exist, picks one, and the MCP server handles the details.
Why it matters
The gap grows with every tool.
Custom connections multiply. MCP plugs only add up. Hover the chart, or change how many AI apps you use.
Custom cablesWith MCP
Simplified: assumes every AI app and tool supports MCP.
Under the hood, step by step: build a tiny MCP server that lets an AI read sales and post in Slack, safely.
Build along
Build a tiny MCP server, one step at a time
We start with one tool, send the assistant the same request each time, and add one small thing per step. Press Run it after each step and watch the result change.
Plug in carefully
Every server you add is code you trust.
MCP makes connecting easy. That includes connecting the wrong thing.
Install only servers you trustTreat them like any dependency: review the source, pin versions.
Give each server the smallest keyA sales reader doesn't need write access to your sales app.
Read the tool descriptionsThey are text the AI reads, so a bad server could hide instructions there.
Watch what gets calledLog tool calls, and let a person approve anything that sends data out. MCP's spec recommends both.3
"Tools"· Model Context Protocol specification, revision 2026-07-28. The tools/list and tools/call messages, input checks and access control on servers, and a person able to deny tool calls.
"MCP Python SDK"· the official Python SDK. Its quickstart: an MCPServer, a function marked @mcp.tool(), and a docstring that describes it.
The chat, the screens, the sales number and the connector counts are made up to show the idea. The build-along output is simulated. Charts marked as illustrations use made-up numbers.