@wukonglabs/wukong

CI-First AI Agent CLI — a minimal yet full-featured AI coding assistant

Wukong is a CLI AI coding agent with a ReAct reasoning loop, built-in tools, a skill system, and session persistence. Powered by @wukonglabs/wukong-agent-sdk.

Install

npm install -g @wukonglabs/wukong@latest

Requires Node.js >= 22. Pure JS/TS, no native addons — install and run, no C++ toolchain needed.

First Run (Interactive Setup)

Run wukong directly. If no model provider is configured, an interactive wizard starts:

🔧 No model config detected — starting interactive setup
Choose a provider type:
  1) OpenAI (Official)
  2) Anthropic (Claude Official)
  3) OpenAI-Compatible (custom baseUrl, e.g. DeepSeek / vLLM)
Enter choice [1]: 3
API Key: sk-xxx
Base URL (required, e.g. https://api.deepseek.com/v1): https://api.deepseek.com/v1
Model name: deepseek-chat
✅ Config saved to /Users/you/.wukong/wukong.json

Config is written to ~/.wukong/wukong.json (permissions 0600). You can also manually configure as described below.

Non-interactive environments (pipes / CI) skip the wizard; use env vars or a config file instead.

Usage

# Interactive mode
wukong

# Single-shot
wukong agent --message "Check my package.json"

# JSON output (for programmatic use)
wukong agent --json --message "List files in current directory"

# Reuse session (multi-turn)
wukong agent --message "Continue the last task" --session-id "my_session"

# Custom system prompt
wukong agent --message "Review this PR" --system-prompt "You are a code reviewer"

# Image recognition (requires vision config)
wukong agent --message "Describe this image" --images https://example.com/shot.png

CLI Options

| Option | Short | Description | | ----------------- | ----- | ---------------------------------------- | | agent | - | Subcommand: single-shot mode | | --message | -m | Message content | | --session-id | -s | Session ID (default: cli:local) | | --system-prompt | - | Injected system prompt | | --images | - | Image URLs (multiple allowed) | | --json | - | JSON output | | --test | - | Test mode: send test message then exit |

Configuration

Priority: CLI args > env vars (.env) > config file (wukong.json) > defaults

Config file lookup order: $WUKONG_CONFIG_PATH → ./wukong.json → ~/.wukong/wukong.json

Minimal Config File

Place at ~/.wukong/wukong.json (global) or ./wukong.json (project):

{
  "agent": {
    "defaultModel": "deepseek-chat"
  },
  "llm": {
    "providers": {
      "deepseek": {
        "type": "openai-compatible",
        "apiKey": "${DEEPSEEK_API_KEY}",
        "baseUrl": "https://api.deepseek.com/v1"
      }
    },
    "models": {
      "deepseek-chat": {
        "provider": "deepseek",
        "id": "deepseek-chat"
      }
    }
  }
}

Use --model <alias> at startup. In TUI, /model lists models and /model <alias> switches between turns.

Environment Variables (Recommended)

OPENAI_API_KEY=sk-xxx              # → llm.providers.openai.apiKey (if declared)
LLM_BASE_URL=https://api.xxx/v1    # → llm.providers.openai.baseUrl
LLM_MODEL=openai-default           # → agent.defaultModel (declared alias)
ANTHROPIC_API_KEY=sk-ant-xxx       # → llm.providers.anthropic.apiKey (if declared)

Config file string values support ${VAR_NAME} templates referencing env vars.

Full config options (multi-provider, Vision, Skills) → root README.

Skills

Inject domain knowledge and custom tools via SKILL.md files — no source changes needed. Skills auto-load from <install-dir>/skills/bundled/, ~/.wukong/skills/, <project>/skills/, and configured directories (later paths override earlier ones by name). See root README for details.

License

MIT