What Is MCP
Model Context Protocol (MCP) is a standard protocol for AI agents to connect to external tools/data sources. Proposed by Anthropic in 2024, with many clients/servers published from 2025.
Why a Standard Was Needed
Each LLM vendor had its own Tool Use, Function Calling spec, requiring per-vendor implementation per tool. MCP is, like USB-C, a standard where "make once, use on multiple clients."
3 Kinds of Capability
Tools
Functions the LLM can call. E.g., "SELECT on a database," "file write," "external API call."
Resources
Data the LLM can read. E.g., "file list," "latest log," "DB schema."
Prompts
Reusable prompt templates. E.g., "code-review checklist," "minutes template."
Server Implementation (TypeScript)
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
const server = new Server(
{ name: "weather-server", version: "1.0.0" },
{ capabilities: { tools: {} } }
);
server.setRequestHandler("tools/list", async () => ({
tools: [
{
name: "get_weather",
description: "Get the weather",
inputSchema: {
type: "object",
properties: { city: { type: "string" } },
required: ["city"],
},
},
],
}));
server.setRequestHandler("tools/call", async (req) => {
if (req.params.name === "get_weather") {
const city = req.params.arguments.city;
const weather = await fetchWeather(city);
return { content: [{ type: "text", text: weather }] };
}
throw new Error("Unknown tool");
});
const transport = new StdioServerTransport();
await server.connect(transport);Server Implementation (Python)
from mcp.server import Server
from mcp.server.stdio import stdio_server
server = Server("weather-server")
@server.list_tools()
async def list_tools():
return [{"name": "get_weather", "description": "...", "inputSchema": {...}}]
@server.call_tool()
async def call_tool(name, arguments):
if name == "get_weather":
return [{"type": "text", "text": fetch_weather(arguments["city"])}]
async def main():
async with stdio_server() as (read, write):
await server.run(read, write, ...)
Client Configuration
Example Claude Desktop config file:
{
"mcpServers": {
"weather": {
"command": "node",
"args": ["/path/to/weather-server/index.js"]
}
}
}Main MCP Servers (Official/Community)
- filesystem: file read/write
- github: Issue / PR operations
- slack: message send/search
- postgres: SQL execution
- puppeteer: browser operation
- memory: cross-conversation memory
- sequential-thinking: thinking-process support
- browserbase: cloud browser
Security Considerations
- Always implement permission checks on the server side
- Restrict access to confidential files (filesystem server uses ALLOWED_DIRS env var)
- When using a third-party MCP server, always read the code
- Mind supply-chain attacks (malicious servers via npm/PyPI)
Production Operation Patterns
Local (stdio)
Claude Desktop local execution. Launched as a child process.
Remote (HTTP / SSE)
The HTTP+SSE transport added in 2025. Can be provided to multiple users via cloud. Deployable to Cloudflare Workers etc.
Testing and Debugging
- Visualize requests/responses with the MCP Inspector (official tool)
- Unit test: call tool functions directly
- Integration test: call from an actual client (Claude Desktop)
2026 Trends
- OpenAI, Google moving to MCP support
- Marketplaces (Smithery, Glama) gaining momentum
- Standardization of Auth, streaming responses
- Feature for agents to dynamically discover/connect MCP
Summary
MCP is the "USB-C" of the agent era. With the ease of writing in ~100 lines of TypeScript/Python and the distribution power of being usable immediately from Claude Desktop/Code, it spread explosively in 2025-2026. Integrating with internal tools raises AI agents' practicality a notch.
