MCPShell
A tool that lets AI agents safely run shell commands, CLI tools and scripts as MCP tools, with YAML-defined tool configuration and CEL-based parameter validation.
MCPShell lets an LLM safely execute command-line tools through the Model Context Protocol, acting as a secure bridge between the model and operating-system commands. Tools are defined in YAML configuration files with parameters, constraints and output formatting, rather than in code, so a new tool can be prototyped by adding a small shell command.
Parameter values are validated with CEL expressions before a command runs, and commands can optionally run inside sandboxed environments for additional isolation. It works with any MCP-compliant client, including Cursor, VS Code and Witsy, and configuration examples cover use cases such as disk-space troubleshooting, Kubernetes with kubectl, and AWS CLI commands.
It is written in Go, runs directly with go run, referencing a versioned module path plus a tools directory and log file, and can also be deployed in containers or Kubernetes.