mcp-use — Fullstack MCP Framework for AI Agents & Apps

What it is

mcp-use is a fullstack MCP (Model Context Protocol) framework for building servers and applications that work with ChatGPT, Claude, and any LLM that supports MCP. It provides both TypeScript and Python SDKs, an interactive inspector for debugging, cloud deployment tooling, and UI widgets for building agent interfaces.

The framework targets developers building MCP-powered tools who want a structured approach rather than wiring raw protocol handlers manually. It handles transport, serialization, tool registration, and client-side rendering so you can focus on the business logic of your MCP tools.

How it saves time or tokens

Building an MCP server from scratch requires understanding the protocol specification, implementing JSON-RPC transport, handling tool schemas, and managing lifecycle events. mcp-use abstracts these layers into a declarative SDK where you define tools as functions and the framework handles everything else. The inspector lets you test tools interactively before connecting to an LLM, catching issues early. Cloud deploy packages your server as a container with a single command.

How to use

1. Create a new MCP app with TypeScript:

npx create-mcp-use-app@latest

Or install the Python SDK:

pip install mcp-use

1. Define a tool in your server:

import { McpServer } from 'mcp-use';

const server = new McpServer({ name: 'my-tools' });

server.tool('get_weather', {
  description: 'Get current weather for a city',
  parameters: { city: { type: 'string' } },
  handler: async ({ city }) => {
    const data = await fetchWeather(city);
    return { temperature: data.temp, condition: data.condition };
  }
});

server.start();

1. Test with the inspector, then add to your AI assistant's MCP config.

Example

Python SDK example:

from mcp_use import McpServer

server = McpServer(name='my-tools')

@server.tool(description='Search documents by query')
def search_docs(query: str, limit: int = 10):
    results = db.search(query, limit=limit)
    return [{'title': r.title, 'snippet': r.snippet} for r in results]

server.start()

Related on TokRepo

• MCP Chrome Integration — browser automation through MCP
• MCP GitHub Integration — GitHub access via MCP protocol

Common pitfalls

• The TypeScript and Python SDKs have slightly different API surfaces; do not assume code translates directly between them
• Cloud deploy requires Docker; ensure your tool dependencies are included in the container build
• Interactive widgets are client-side only and do not work in headless MCP connections from CLI tools like Claude Code