AX CLI Features (Current)

January 20, 2026 · View on GitHub

Last reviewed: 2025-02-21 Applies to: ax-cli/ax-grok v4.4.x

This document lists the capabilities that are actually supported today. It replaces older marketing-style writeups.

Note: The ax-glm package has been deprecated. GLM/Z.AI users should use OpenCode - the official CLI from Z.AI.

Providers and defaults

  • ax-grok (xAI): default model grok-4 (alias to latest 4.x). Reasoning/search/vision built in. Variants include grok-4.1-fast-reasoning, grok-4.1-fast-non-reasoning, grok-4.1-mini, grok-4-0709, plus grok-2-image-1212 for images.
  • ax-cli (local/offline): defaults to Ollama http://localhost:11434/v1 with qwen3:14b starter model. Supports other local models (qwen3 family, qwen2.5-coder:32b, codegeex4).

Core experience

  • Interactive and headless modes with tool execution (file edits, search, bash).
  • Thinking/reasoning: Grok uses reasoning_effort where supported; toggled via --think/--no-think or settings.
  • Session resume (--continue) per working directory with isolated state per provider (.ax-grok, .ax-cli).
  • Structured output for editor/automation via --json and VSCode-friendly formatting via --vscode.
  • Deterministic mode (--deterministic/--seed) and sampling control (--top-p, temperature).

Project initialization (/init)

  • Generates AX.md at project root - a single-file AI context document.
  • Four depth levels: basic (quick scan), standard (default), full (architecture), security (audit).
  • Adaptive output: verbosity scales with project complexity (file count, LOC, dependencies).
  • Parses existing rules from .cursorrules, .editorconfig, and similar files.
  • Complexity scoring categorizes projects as small/medium/large/enterprise.
  • Deep analysis mode (full/security) generates .ax/analysis.json with dependency graphs and hotspots.
  • Migration support: replaces legacy 3-file format (CUSTOM.md, ax.index.json, ax.summary.json).

Automatic context injection

  • When starting a conversation, AX.md is automatically read and injected into the AI's system prompt.
  • Content is wrapped in <project-context source="AX.md"> tags for clear delineation.
  • HTML metadata comments (generation date, etc.) are stripped to save tokens.
  • Priority: AX.md (primary) → ax.summary.json (legacy) → ax.index.json (legacy).
  • Benefits: AI understands your project's build commands, tech stack, and conventions from the start.

Configuration and security

  • Encrypted API key storage created by ax-<provider> setup; env var overrides for CI.
  • Precedence: flags > env (AI_MODEL, AI_BASE_URL, provider key) > project settings (.ax-<provider>/settings.json) > user settings (~/.ax-<provider>/config.json) > defaults.
  • Loop detection for tool calls and configurable tool round limits (see config-defaults/settings.yaml).

Integrations

  • Model Context Protocol (MCP) client support (see docs/mcp.md and templates).
  • VSCode extension integration (status bar, thinking indicator, JSON output).
  • Multi-provider isolation to run multiple CLIs in parallel without state conflicts (docs/multi-provider-guide.md).

What is not supported

  • No guarantee of offline operation unless you use local backends (Ollama/LM Studio/vLLM) with ax-cli.