MCP, the Model Context Protocol, is an open protocol published by Anthropic in November 2024 under the Apache 2.0 license. It standardizes how AI applications connect to external tools, data sources, and prompt templates through a client and server interface built on JSON-RPC 2.0. The headline value: implement the spec once on either side and any compliant host can talk to any compliant server, collapsing M times N custom integrations into M plus N.
**The Model Context Protocol (MCP)** is an open standard published by Anthropic in November 2024 that defines a single wire format for AI applications to talk to external tools, data sources, and prompt templates. It is licensed under Apache 2.0 so any vendor can adopt it freely.
The analogy that sticks: MCP is the USB-C of AI integrations. Before MCP, every AI host like Claude Desktop or Cursor had to write a custom adapter for every external service it wanted to reach. With MCP, the host implements the protocol once and gains access to every MCP server in the ecosystem.
Underneath, MCP is built on JSON-RPC 2.0, a well established message format. Each message is a small JSON object with a method name and parameters. The protocol defines specific methods for listing tools, calling tools, reading resources, and fetching prompts. The transport layer (how the bytes actually move) is separate: a local server can run as a subprocess over standard input and output, while a remote server can run as an HTTP service.
What the ecosystem looks like in 2026: official SDKs in Python, TypeScript, Kotlin, Java, Rust, C#, and Swift, plus thousands of community built servers covering GitHub, Slack, Postgres, your local filesystem, and most popular SaaS tools.
At a meetup, here is how to explain it. **The Model Context Protocol (MCP)** is an open, Apache 2.0 licensed protocol that Anthropic introduced in November 2024 to standardize how AI applications connect to external context. By 2026 it is governed under a broad open ecosystem with official SDKs in Python, TypeScript, Kotlin, Java, Rust, C#, and Swift.
The protocol defines a client and server interface built on JSON-RPC 2.0. A host application (Claude Desktop, Cursor, Zed, Claude Code, the OpenAI Agents SDK) embeds an MCP client. That client opens a connection to an MCP server, which exposes three primitive families: tools the model can call, resources the host can pull for context, and prompt templates the user can invoke.
The headline value is the network effect. Before MCP, every host wrote a bespoke integration for every tool and every tool vendor wrote a bespoke integration for every host. The cost was M times N. After MCP, each side implements the protocol once and the integration cost collapses to M plus N. Implement the spec on either side and any compliant host can talk to any compliant server.
The protocol is intentionally small. It covers initialization with capability negotiation, tool calls, resource reads, and prompt renders, plus helpers for completion, sampling, and roots. Everything outside that scope is left to the server or host. This narrow surface is why adoption moved fast across Claude Desktop, Cursor, Zed, Continue, Sourcegraph Cody, Windsurf, and Gemini CLI.
Quick reference
TL;DR
MCP is an open Apache 2.0 protocol from Anthropic (November 2024) that standardizes how AI apps connect to tools, data, and prompts via JSON-RPC 2.0, turning M times N integrations into M plus N.
Imagine every coffee machine needing a different plug shape for the wall socket. Every kitchen would need a drawer full of adapters. MCP is the agreement that says all coffee machines (AI apps like Claude Desktop or Cursor) and all wall sockets (data sources like GitHub, your filesystem, or Postgres) use the same plug shape. The plug was designed by Anthropic in November 2024 and given away under an open license so any company can adopt it freely. Once a tool ships an MCP server, every MCP aware AI app can use it. Once an AI app speaks MCP, it can reach every existing server. The drawer of adapters disappears, and adding a new coffee machine or a new socket costs nothing extra on the other side.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
A one off internal script that only ever talks to one private API MCP shines when many hosts meet many tools. For a single closed pairing, a direct SDK call is simpler and adds no protocol overhead.
An offline batch job with no model in the loop MCP exists to connect an LLM host to external context. If no host or model is involved, the protocol adds complexity with no value.
Ultra low latency pipelines where every microsecond counts JSON-RPC over stdio or HTTP adds serialization and framing overhead. Tight numerical loops belong in compiled code, not behind a protocol boundary.
Quick scan · 2–4 min
Everything important, quickly.
6 min: give the one sentence definition, name Anthropic and November 2024, explain JSON-RPC 2.0 as the substrate and Apache 2.0 as the license, walk through the M plus N versus M times N economics, and name three hosts and three servers shipping in 2026.
Real-world examples
Real products, models, and research that use this idea.
1Claude Desktop ships with an MCP client and pre installed filesystem, fetch, and memory servers; users add more from the public registry by editing the config file.
2Cursor and Zed both embed MCP clients in 2026, so the same GitHub or Postgres server works across both editors without any host specific glue code.
3The OpenAI Agents SDK released MCP adapter support in 2025, letting GPT based agents reach the same server ecosystem Claude uses.
4Google Gemini CLI added native MCP client support in 2026, demonstrating the protocol is vendor neutral in practice.
5Thousands of community servers are indexed at the official servers repo, glama.ai, mcp.so, and smithery.ai.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy JSON-RPC 2.0 and not gRPC or a custom binary protocol?
Sign in or go Pro to see the approach for this follow-up.
QHow does MCP relate to function calling on the LLM side?
Sign in or go Pro to see the approach for this follow-up.
QWhat stops a malicious MCP server from doing damage on the host machine?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Calling MCP a library or an SDK. MCP is a wire protocol that defines messages on a transport. The SDKs in Python, TypeScript, and other languages are convenience implementations, not the protocol itself.