MCP.run is a registry and runtime platform for hosting and executing MCP (Model Context Protocol) tools using WebAssembly-based isolation, built by Dylibso to provide secure, portable, and universal tool execution for AI applications. It solves the challenge of safely running third-party MCP servers by executing tools as WebAssembly modules (called servlets) in a sandboxed environment, eliminating the need for heavy containers or virtual machines while providing strong security isolation. MCP.run enables anyone to publish, discover, and execute MCP tools that work across any platform, operating system, processor, or device.
MCP.run differentiates itself with its WebAssembly-first architecture where all servlets are compiled to Wasm modules, ensuring portability and security through strict sandboxing. The platform requires developers to explicitly declare domain allowlists, environment variable access, and file system paths ahead of time, preventing unauthorized network access or data exfiltration. MCP.run provides client connectors including mcpx for native integration, mcpx4j for Java and Android support, and web-based interfaces for browsing and testing available tools. The registry allows anyone to publish servlets with transparent security policies that users can review before execution.
MCP.run is designed for AI developers, platform teams, and organizations that need to integrate third-party MCP tools into their AI applications with production-grade security and minimal operational overhead. It integrates with any MCP-compatible AI client including Claude, ChatGPT, Cursor, and custom agent frameworks, providing a centralized hub for discovering and executing tools without managing individual server deployments. The platform is particularly valuable for enterprise environments where security policies require strict control over what external tools can access, offering auditable, sandboxed execution that meets compliance requirements while maintaining the flexibility of the MCP ecosystem.