mcp-use – TypeScript Framework for MCP Servers and Apps

TL;DR

TL;DR: mcp-use is an MIT-licensed TypeScript and Python framework for building MCP servers, ChatGPT plugins, and AI apps — with Zod schemas, React Views, and a built-in Inspector for debugging.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: mcp-use.com ← MUST visit and verify
• Source repository: github.com/mcp-use/mcp-use ← MUST read README
• License: MIT (verified via GitHub license.spdx_id)
• HN launch thread: news.ycombinator.com/item?id=48762862 — Manufact (YC S25)
• Manufact (builder company): manufact.com
• Source last checked: 2026-08-24 (commit main, pushed 2026-08-23)

What Is mcp-use?

mcp-use is the open-source SDK built and maintained by Manufact (YC S25) for building applications powered by the Model Context Protocol. The README describes it as:

“The TypeScript framework for MCP. Build, test, and ship MCP servers, ChatGPT plugins, Claude connectors.”

It ships in two language flavors — TypeScript and Python — and supports both the classic MCP protocol and the MCP 2026 protocol revision. The TypeScript SDK is the reference implementation and gets new features first.

Key capabilities (directly from the README):

• Fully typed — Zod schemas flow from tool definitions through to structured results, View props, and tool calls
• Native Views — bind React Views directly to tools and ship interactive apps without custom extension wiring
• Agent-first and headless — scaffold, invoke, inspect, screenshot, and deploy through your agent
• Built-in debugging — Inspector tool runs in the browser or headlessly via CLI
• Built-in tunneling — expose local HTTP, WebSocket, and MCP servers through a managed relay
• One-line OAuth adapters — drop in OAuth without building provider wiring from scratch

Ecosystem Packages

The monorepo publishes several packages (verified from the README Ecosystem table):

| Package | Use it for | |---|---| | mcp-use | TypeScript v2 server framework, React Views, and CLI | | @mcp-use/client | Connect to MCP servers from Node.js, browsers, React, and sandboxes | | @mcp-use/agent | Build model-powered agents on top of MCP clients | | @mcp-use/inspector | Inspect and debug MCP servers and apps | | @mcp-use/tunnel | Expose local HTTP, WebSocket, and MCP servers through the managed relay | | create-mcp-use-app | Scaffold servers and interactive apps | | mcp-use for Python | Build Python MCP servers, clients, and agents |

Setup Workflow

Step 1: Scaffold a project

The fastest path in is the official scaffold:

npx -y create-mcp-use-app@latest

This generates a complete project with TypeScript configuration, development scripts, the Inspector, and a React view pipeline.

Step 2: Run locally

npm run dev

The server starts and the MCP Inspector becomes available at http://localhost:3000/mcp/inspector.

Step 3: Define a tool with Zod

Replace index.ts with a view-bound tool:

import { MCPServer } from "mcp-use";
import { z } from "zod";

const server = new MCPServer({
  name: "weather-app",
  title: "Weather App",
  version: "1.0.0",
});

server.addTool({
  name: "get-weather",
  description: "Fetch weather for a city",
  schema: z.object({
    city: z.string().describe("City name"),
  }),
  handler: async ({ city }) => {
    // Return structured data
    return { temp: 22, condition: "sunny" };
  },
});

await server.run();

Step 4: Inspect and debug

Open http://localhost:3000/mcp/inspector to see all registered tools, invoke them with sample inputs, and observe responses — all without a separate package install.

Step 5: Deploy

npx mcp-use deploy

Or self-host via Docker (MIT licensed, docker pull the image from GitHub Container Registry — check the README for the exact image name and tag).

Deeper Analysis

Why mcp-use instead of FastMCP or the official Anthropic SDK?

mcp-use occupies a broader niche than FastMCP (which is server-only) or the official Anthropic SDK. Its distinguishing characteristics:

Cross-client compatibility. The same server works with ChatGPT, Claude, and any other MCP-compatible host. FastMCP is tied to the Anthropic client ecosystem.

Native Views on MCP 2026. The README benchmark table shows nativeViewsMcp: ✅ as a key differentiator — native Views on the MCP 2026 protocol, which the official SDK does not support.

TypeScript-first with full parity to Python. The Python package (pip install mcp-use) is a first-class citizen with protocol conformance badges in the README, not a rushed port.

Built-in Inspector and tunneling. No need to install a separate inspector package or configure ngrok — both are included in the CLI.

Performance

The README benchmarks section shows mcp-use achieving 10,982 ops/s versus FastMCP’s 6,628 ops/s and the official SDK’s 8,050 ops/s. Install size is also the smallest of the group at 74.4 MiB for a dev stack versus 122+ MiB for the alternatives.

Manufact Cloud

Manufact Cloud is the hosted deployment platform from the same team. It provides managed infrastructure for MCP servers built with mcp-use. The source SDK is MIT-licensed regardless of whether you self-host or use Manufact Cloud.

Practical Evaluation Checklist

• Install the scaffold and run npm run dev — Inspector loads without extra steps
• Verify tool schema validation by passing invalid input and observing Zod errors
• Test the built-in tunneling with @mcp-use/tunnel from the CLI
• Connect the server to a ChatGPT or Claude session using the connection URL from the Inspector
• Deploy to Manufact Cloud with npx mcp-use deploy
• Run mcp-use Python variant: pip install mcp-use and scaffold with the Python equivalent

Security Notes

• The repository has a SECURITY.md policy — review it before submitting sensitive tool handlers
• When tunneling to the managed relay, be aware that traffic passes through Manufact’s infrastructure; do not tunnel production databases without VPN wrapping
• OAuth adapters should be reviewed for token storage practices; prefer short-lived tokens with refresh flows

FAQ

Q: Is mcp-use tied to a specific LLM provider? A: No. It implements the Model Context Protocol specification, making it compatible with any MCP host — ChatGPT, Claude, and others that adopt MCP.

Q: Is the Python package as fully featured as the TypeScript version? A: The Python SDK covers servers, clients, and agents with protocol conformance badges. The TypeScript SDK is the reference implementation and gets new features first.

Q: How does mcp-use differ from the official Anthropic MCP SDK? A: The official Anthropic SDK is focused on building clients and servers for Anthropic products. mcp-use is cross-client, includes React Views, a built-in Inspector, tunneling, and benchmarks faster than alternatives.

Q: Can I self-host an mcp-use server? A: Yes. The SDK is MIT-licensed and the README documents Docker-based self-hosting. Manufact Cloud is optional.

Q: What is the Manufact company behind this? A: Manufact is a Y Combinator S25 company. They maintain mcp-use and offer Manufact Cloud for hosted deployments.

Conclusion

mcp-use is a production-ready MCP framework with 10k+ GitHub stars, MIT licensing, and active maintenance by a YC-backed team. Its TypeScript-first design, built-in Inspector, native Views, and cross-client support make it one of the most complete MCP development platforms available. The Python SDK brings the same capabilities to non-TypeScript codebases.

Quick-start: npx -y create-mcp-use-app@latest, then npm run dev to open the Inspector.

• Source: github.com/mcp-use/mcp-use
• Docs: docs.mcp-use.com
• Live Inspector: inspector.mcp-use.com