mcp.run Review 2026

Serverless platform for running and composing MCP servers (called 'servlets') in a portable WebAssembly sandbox, with a marketplace for installing tools into any MCP client.

mcp.run from Dylibso applies the WebAssembly-component model to the Model Context Protocol: instead of installing MCP servers as Node or Python processes, developers package them as portable Wasm 'servlets' that the platform can execute in a hardened sandbox on demand. Each servlet declares the tools it exposes, the host capabilities it needs (network, filesystem, env), and a manifest the runtime enforces — eliminating a whole class of supply-chain risks that come with running arbitrary Node packages. The mcp.run registry lets users browse and install servlets into Claude Desktop, Cursor, Cline, Goose, or any MCP-compatible client through a small local proxy. Because servlets are language-agnostic Wasm components, the same tool can be authored in Rust, Go, TypeScript, or Python and run identically across machines, making it especially appealing for teams that need reproducible, auditable agent tooling. The platform offers private workspaces, secret management, usage logs, and the ability to publish proprietary servlets to a team. For builders thinking about long-term agent infrastructure, mcp.run represents a credible answer to 'how do we let agents run third-party code without trusting it entirely.' Pricing details should be verified on the live site as plans evolve.