mcp-use

What is mcp-use?

mcp-use is the open-source developer toolkit and cloud infrastructure purpose-built for the Model Context Protocol (MCP). Co-founded by Luigi and Pietro Zullo (YC S25), it helps engineering teams quickly build and ship custom AI agents that connect to real-world services through MCP servers.

The project started from a simple frustration: when MCP first launched, it could only be used inside IDEs like Cursor or Claude Desktop. mcp-use was created so that developers could write agents in code, configure MCP servers programmatically, and run them safely in production — not just in a local sandbox.

Who uses it?

The SDK has surpassed 6,000 GitHub stars and 120,000 downloads, with adoption at companies ranging from startups to large enterprises including NASA, NVIDIA, Cisco, and SAP.

How it works — 3-layer architecture

• mcp-use SDK — Open-source Python/TypeScript library. Lets you write agents in code, configure multiple MCP servers into a single unified pool, and run them locally or in the cloud.
• mcp-use Cloud Platform — The central control plane. Manages server configs, tool selection, caching, metrics, access control, and authentication. Think of it as an MCP gateway all requests pass through.
• mcp-use Server Hosting — Spin up managed or self-hosted MCP servers, including short-lived stdio sandboxed servers for safe execution.

Key problems it solves

• Fragmented MCP server configs scattered across GitHub repos, hardcoded values, and random registries
• No standard way to handle auth, access control, or audit logging for MCP calls
• Tool overload — exposing too many tools to an LLM causes confusion and degraded performance
• Agents running locally with no observability when something breaks at 2am
• No safe, observable runtime for production MCP agents

Key features

• Unified multi-server pooling — configure and manage multiple MCP servers from one interface
• Profile-based access control — restrict which tools and servers each agent or user can access
• Audit logs and observability — full traceability of every tool call, payload, and server response
• Tool limiting — expose only the tools you need to keep LLMs focused
• Sandboxed agent runtime — safe execution environment for production workloads
• Environment management — separate dev, staging, and production MCP configurations

The team describes mcp-use as “Vercel + Next.js, but for MCP” — a complete vertical solution from open-source SDK to production-ready cloud infrastructure.