Grok CLI Open source

What it does

grok-cli is an open-source terminal coding agent built specifically around xAI's Grok API. Beyond the standard "read code, edit files, run commands" agent loop, it adds live X/web search, foreground and background sub-agent delegation, a built-in desktop-automation sub-agent, and remote control of a running session from Telegram.

What is inside

Foreground task and background delegate sub-agents, a /verify flow that builds/runs/tests an app in a sandbox with screenshot/video evidence, image and video generation tools, a computer sub-agent for macOS desktop automation (via agent-desktop), MCP server support, Agent Skills (.agents/skills/<name>/SKILL.md), lifecycle hooks, and an optional microVM sandbox (macOS 14+ Apple Silicon only) that isolates shell commands from the host.

Works with

xAI's Grok models (grok-4.3, grok-4.20 variants, etc.); MCP servers; Telegram for remote pairing; a companion tool for macOS accessibility automation.

How to install or connect

curl -fsSL https://raw.githubusercontent.com/superagent-ai/grok-cli/main/install.sh | bash

Alternative: bun add -g grok-dev.

Maintenance and safety

MIT licensed, built by superagent-ai with CI and clear docs. The main branch's most recent commit is from May 2026. Desktop automation and Telegram remote control meaningfully widen the attack surface: the README requires granting terminal-app Accessibility permissions on macOS and explicitly warns to treat Telegram bot tokens like a password. The sandbox only covers Apple Silicon, so other platforms run agent shell commands unsandboxed.

Who should use it

Developers already using Grok/xAI who want a hackable, fully open-source terminal agent with search, media generation and phone-based remote control, and who are comfortable managing the extra credentials (Grok API key, Telegram token) its extra features require. It's also a reasonable pick for anyone specifically wanting to kick off and check on long-running coding tasks from a phone via Telegram rather than staying tied to a laptop, or who wants live X/web search built into the agent loop instead of bolting on a separate MCP search server.