UI-TARS Commands MCP

ByteDance's Commands MCP server, from its UI-TARS Desktop repository and published on npm as @agent-infra/mcp-server-commands. It gives your agent a shell: it runs a command or script on the machine where the server runs and sends back the output. ByteDance adapted it from g0t4's mcp-server-commands. What you can do with it - Run a command with its arguments, such as git status or ls -la, and get back what it printed and any errors (run_command) - Pipe a script to an interpreter such as bash, zsh, fish, python or node and get its output (run_script) - Choose the working directory for each call, or set a default with --cwd - Use its run_command prompt to put a command's output into your conversation, where you pick the command rather than the agent How it connects It runs on your machine with npx @agent-infra/mcp-server-commands@latest and needs Node.js 18 or newer. ByteDance documents setup for VS Code with GitHub Copilot, Cursor, Windsurf and Claude Desktop. Start it with --port to serve it over Streamable HTTP (/mcp) or SSE (/sse) on 127.0.0.1 instead of stdio. The risk - It runs whatever command or script your agent sends, as your user account: it can read, change and delete your files, install software and reach the network. - There's no allowlist, sandbox or read-only mode. --cwd only sets where commands start; it doesn't keep them inside that folder. - Your agent's tool approval is the only check, so read each command before you allow it. - It listens on 127.0.0.1 by default. Setting --host 0.0.0.0 lets anyone who can reach that port on your network run commands on your machine.