Grok CLI

An agentic coding harness for xAI's Grok models. It reads your code, edits it, runs your tests, and iterates — in a terminal, with permission controls you set.

Nothing runs a command or edits a file without your say-so, unless you say so once.

Approval prompts for edits show a diff , so you approve a change rather than a filename. "Always allow" is scoped to the tool and its argument — approving cargo test never approves rm -rf / .

Rules are Tool(pattern) . cmd:* is a prefix match, anything else is a glob, and a bare Tool matches every call. Deny always wins — including over bypassPermissions , and including inside subagents.

Shell commands are matched whole and per segment, so true; rm -rf / cannot slip past a rule by not starting with the denied text. Matching is still textual, though: rm -fr is not rm -rf , and a command built from a variable cannot be inspected. Treat deny rules as a guardrail against mistakes, not a sandbox — plan mode with an explicit allow-list is the stronger control.

Hooks — run a program before or after any tool call. It can veto.

Subagents — .grok/agents/reviewer.md , for work that would otherwise fill the conversation with intermediate output:

A subagent's tools are always a subset of the parent's, so delegating can never escape a restriction.