Anthropic released the Model Context Protocol in November 2024. By March 2026, every major AI company — OpenAI, Google, Microsoft, Amazon — has adopted it. The official registry has over 6,400 MCP servers. Monthly SDK downloads crossed 97 million.
That's not hype. That's infrastructure.
The Problem MCP Kills
Before MCP, connecting an AI model to external tools was an N×M nightmare. Every model needed custom integration code for every tool. Want Claude to read your database? Write a connector. Want GPT-4 to access your CRM? Write another one. Want Gemini to query your file system? Yet another.
N models times M tools equals a pile of bespoke integrations that nobody wants to maintain.
MCP flips this to N+M. Build one MCP server for your tool, and every MCP-compatible model can use it. Build one MCP client into your model, and it can talk to every MCP server. Same logic as USB: one standard port, universal compatibility.
How It Works (Without the Jargon)
MCP uses a client-server architecture with three roles:
- Hosts — the AI application (Claude Desktop, an IDE, your custom app)
- Clients — maintain connections between hosts and servers
- Servers — expose tools, data, and capabilities
Servers offer three primitives:
- Tools — functions the AI can call (query a database, send an email, run code)
- Resources — data the AI can read (files, API responses, live feeds)
- Prompts — reusable templates for common tasks
Transport happens over stdio (local) or HTTP with Server-Sent Events (remote). The latest spec adds OAuth 2.1 for authentication and built-in human-in-the-loop approval for sensitive operations.
Why It Won
MCP won because Anthropic made two smart moves. First, they open-sourced it from day one. No proprietary lock-in, no licensing games. Second, they shipped working SDKs in Python and TypeScript before anyone asked.
When OpenAI adopted MCP in early 2025, it was over. The "which standard will win" debate ended before it started. Google followed. Then Microsoft built it into VS Code and Copilot. Amazon integrated it into Bedrock.
The network effect is brutal: more servers make the protocol more valuable, which attracts more clients, which attracts more servers.
MCP vs Google's A2A
Google's Agent-to-Agent (A2A) protocol launched as a complement, not a competitor. MCP connects models to tools. A2A connects agents to other agents. Think of MCP as giving an agent hands to use tools, and A2A as giving agents the ability to collaborate with each other.
In practice, you'll use both. An agent uses MCP to access databases and APIs, then uses A2A to hand off subtasks to specialized agents.
Building Your First MCP Server
It takes about 30 minutes if you already know TypeScript or Python:
npm install @modelcontextprotocol/sdk
Define your tools, resources, and prompts. Implement the handlers. Run it. Claude, GPT, Gemini, and anything else with an MCP client can now use your server.
The official docs at modelcontextprotocol.io are genuinely good — clear examples, working code, no filler.
Where This Goes
MCP is still early despite the adoption numbers. The security model needs hardening. Discovery and trust mechanisms for remote servers are still evolving. Performance optimization for high-throughput use cases is ongoing.
But the trajectory is clear. MCP is becoming the default way AI systems interact with the outside world. If you build tools, APIs, or platforms, supporting MCP isn't optional anymore — it's table stakes.
Top comments (0)