@frontmcp/sdk weekly downloads statistics

<div align="center"> <a href="https://frontmcp.dev"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/agentfront/frontmcp/v1.9.6/.github/assets/readme/hero.svg"> <img src="https://raw.githubusercontent.com/agentfront/frontmcp/v1.9.6/.github/assets/readme/hero.light.svg" alt="FrontMCP - The TypeScript way to build MCP servers" width="100%"> </picture> </a> </div>

@frontmcp/sdk

The core of FrontMCP, a TypeScript framework for building Model Context Protocol (MCP) servers. You write tools, resources and prompts as decorated classes; the SDK handles the protocol, the transports, sessions, authentication, dependency injection and the request pipeline. The same server runs on your machine, in Docker, on Vercel or AWS Lambda, and on Cloudflare Workers.

Use it to build an MCP server, or to run one inside another program. The frontmcp CLI creates, runs, tests and builds the project around it.

Install

Start a new project with the CLI, which sets everything up:

npx frontmcp create my-app

Or add the SDK to a project you already have:

npm install @frontmcp/sdk reflect-metadata
npm install -D frontmcp typescript @types/node@^24
npx frontmcp init

FrontMCP needs Node.js 24 or later. frontmcp init turns on experimentalDecorators and emitDecoratorMetadata in tsconfig.json. Import reflect-metadata once, at the top of your entry file, before anything from FrontMCP.

npm and pnpm install the SDK's required peer dependencies (express, cors, content-type and raw-body) for you; with Yarn, add them yourself. The optional ones turn on features:

Install	To
@frontmcp/storage-sqlite	Keep sessions, tasks and events in SQLite
@frontmcp/observability, @opentelemetry/api, @opentelemetry/sdk-trace-base	Trace and log requests with OpenTelemetry
openai or @anthropic-ai/sdk	Run an @Agent on OpenAI or Anthropic models
@enclave-vm/core	Run dynamic jobs (scripts) in a sandbox

A server in one file

import 'reflect-metadata';

import { App, FrontMcp, Tool, ToolContext, z } from '@frontmcp/sdk';

@Tool({
  name: 'add',
  description: 'Add two numbers',
  inputSchema: { a: z.number(), b: z.number() },
})
class AddTool extends ToolContext {
  async execute(input: { a: number; b: number }) {
    return input.a + input.b;
  }
}

@App({ id: 'calc', name: 'Calculator', tools: [AddTool] })
class CalcApp {}

@FrontMcp({
  info: { name: 'Demo', version: '0.1.0' },
  apps: [CalcApp],
  http: { port: 3000, entryPath: '/mcp' },
})
export default class Server {}

Save it as src/main.ts and run npx frontmcp dev. The server reloads when you edit it, and answers MCP requests at http://localhost:3000/mcp (at / without entryPath). Try it with the MCP Inspector, npx frontmcp inspector, or add it to Claude Code:

claude mcp add --transport http demo http://localhost:3000/mcp

A tool is a class: @Tool describes it to the model, inputSchema validates its input with Zod (the SDK exports z), and execute returns the result. An @App groups tools, resources, prompts and providers, and @FrontMcp is the server. The Quick Start walks through each piece.

What the SDK covers…