When file‑by‑file copy‑pasting stopped developers, one programmer built a zero‑dependency MCP server in pure Python. The tool lets AI clients call local file‑system functions directly, bypassing the need for heavy frameworks. By using only the standard library, the author avoided adding five extra packages and kept the runtime lightweight.

The server implements the MCP spec with two transports: stdio for a single local client and HTTP/SSE for concurrent connections. A single flag flips between modes, letting the same dispatcher serve five clients in under 50 ms on a Windows 11 machine running Python 3.12.6. Security checks resolve every path before any file operation.

Design decisions stem from a real bottleneck: a recursive file search that kept the server alive for minutes. Switching to a shallow default and requiring explicit recursion eliminated the hang, cutting search time from endless to under 30 ms on a typical project folder. The author also hardened the server against path traversal, symlink abuse, and UNC attacks.

With no external dependencies, the project ships as a single file and includes 50 unit tests covering security and functionality. The source lives on GitHub, ready for contributors to extend or adapt it for other AI toolchains. This minimalist approach demonstrates that a well‑defined protocol can replace heavyweight stacks without sacrificing performance or safety.