Hi HN — I built Terminal MCP, a utility that exposes terminal sessions over MCP, primarily to help LLMs interact with CLIs and TUIs during development and debugging.

It’s also useful outside of AI: it provides a general way to run terminal workflows through a controlled interface, with better isolation and observability than a raw shell.

Recent addition: Sandbox mode, based on Anthropic’s work on safe tool execution. It lets you run sessions with explicit controls over what the process can access — by default blocking common locations where credentials live, and optionally restricting filesystem and network access.

Repo: https://github.com/elleryfamilia/terminal-mcp

Feedback welcome.