---

Why wow-agent?

Traditional AI coding loops look like this:

User → copy code → ask AI → edit yourself → ask again → infinite loop

wow-agent breaks the loop and joins your real workflow:

Give a goal → AI plans → tools execute → verify → done

It is not a chat toy — it is an AI pair programmer that works inside real codebases.

---

Why choose wow-agent?

Core difference	Traditional tools	wow-agent
Model freedom	Locked to one vendor/model	Any OpenAI-compatible API + local models (Ollama / LM Studio)
Runtime	Tied to a specific IDE / cloud	Pure terminal, local execution, full data control
Architecture	x86 only	Native ARM64 (tested on Jetson Orin NX, edge-ready)
Safety	AI executes commands directly	Triple safety: network sandbox + egress guard + double confirm + one-click rollback
Extensibility	Closed ecosystem	MCP ecosystem + custom tools + sub-agents

In one sentence: wow-agent gives you model choice, environment control, and safety decisions.

---

Architecture: four-layer design

┌─────────────────────────────────────────────────────┐
│  Layer 1: Agent Brain                               │
│  ├─ task planning · state machine · sub-agents      │
└─────────────────────────────────────────────────────┘
                            ↓
┌─────────────────────────────────────────────────────┐
│  Layer 2: Developer Tools                           │
│  ├─ file CRUD / diff / undo / code index            │
└─────────────────────────────────────────────────────┘
                            ↓
┌─────────────────────────────────────────────────────┐
│  Layer 3: Model Ecosystem                           │
│  ├─ cloud: DeepSeek / Qwen / OpenAI / Gemini / ...  │
│  └─ local: Ollama / LM Studio                       │
└─────────────────────────────────────────────────────┘
                            ↓
┌─────────────────────────────────────────────────────┐
│  Layer 4: Safety System                             │
│  ├─ sandbox · egress guard · confirm · rollback     │
└─────────────────────────────────────────────────────┘

Layer	Capability	Key tech
Agent Brain	task decomposition, state machine, sub-agents	planning → executing → verifying → done/failed
Dev Tools	file ops, diff/undo, code index	AST Python symbol table, incremental index, diff visualization
Model Eco	unified cloud/local access	16+ providers + Ollama/LM Studio, /model fetches lists online
Safety	sandbox / egress guard / confirm / rollback	unshare -n isolation, egress patterns, /undo

---

🚀 Quick Start

Install (requires uv)

# Ubuntu / Debian / WSL
cd wow-agent && curl -LsSf https://astral.sh/uv/install.sh | sh && uv sync && uv run wow

# Windows (PowerShell, experimental)
cd wow-agent; irm https://astral.sh/uv/install.ps1 | iex
$env:Path += ";$env:USERPROFILE\.local\bin"
uv sync; uv run wow

Subsequent launches:

uv run wow        # or .venv/bin/wow

First launch walks you through provider + API key setup, saved to ~/.wow-agent.env (mode 600).

---

✨ Core Features

🤖 Agent Brain

• Task tree + state machine: auto-decomposes multi-step tasks, planning → executing → verifying → done/failed
• Sub-agents: clean-context code research that returns only conclusions; /review three-level code review
• State machine: planning → executing → verifying → done/failed automatic transitions

🛠 Developer Tools

• File operations: read / write / edit / delete / search; code_index builds a Python symbol table in one scan
• Diff & Undo: colored diff visualization, single-step /undo, full-task rollback on failure, snapshots survive restarts
• Code index: AST-parsed Python symbols, incremental updates — query the index before reading files

🤝 Model Ecosystem

Type	Supported
Cloud	DeepSeek / Qwen / Kimi / GLM / SiliconFlow / Doubao / Hunyuan / MiniMax / OpenRouter / Groq / Mistral / Grok / OpenAI / Gemini / OpenCode Zen
Local	Ollama / LM Studio (no API key needed)
Free tier	OpenCode Zen free tier (includes ox alpha x-preview-f-free)

🛡 Safety System

• Network sandbox: unshare -n network namespace isolation, falls back to proxy blackhole
• Egress guard: auto-blocks curl POST / scp / git push / npm publish etc.
• Double confirm: rm -rf /, mkfs, dd, fork bombs, shutdown, chmod -R 777 / require typed confirmation
• One-click rollback: /undo reverts a single step; failed tasks roll back entirely

---

⚙️ Configuration

Config file: ~/.wow-agent.env (mode 600)

WOW_API_KEY=            # API key
WOW_BASE_URL=           # API base URL
WOW_MODEL=              # model id
WOW_SAFE_MODE=1         # 0 disables sandbox
WOW_MAX_ITER=40         # max iterations per task
WOW_AUTO_COMPACT=55000  # auto-compact threshold
WOW_LANGUAGE=en         # en/zh, default en
WOW_MODEL_CTX=128000    # model context length (for estimation)
WOW_CELL_ASPECT=2.0     # terminal cell aspect (logo scaling)
WOW_UPLOAD_GUARD=1      # egress guard switch

MCP servers: ~/.wow-agent/mcp.json

{
  "servers": {
    "fs": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-fs", "/tmp"]}
  }
}

---

💡 Usage

$ wow
Enter auto mode? y/N    # y = skip step-by-step confirms (dangerous/egress still require approval)
> Help me understand this project structure

Interaction basics:

• !<command>: run shell directly without the model (e.g. !git status)
• /: fuzzy command completion, /help lists everything

Command	Description
/model	provider/model wizard; /model <id> quick switch
/status	session status: context, task progress, undoable changes
/compact [hint]	compress history manually (auto-triggers past threshold)
/mem save/use/rm/list/new	long-term memory
/resume	restore a previous session
/undo	revert last change; full rollback on task failure
/review [path]	read-only review: 🔴 high risk / 🟡 suggestion / 🟢 polish
/language [en|zh]	switch language (default en, interactive menu)
/safe	toggle safe mode
/clear /exit	clear conversation / quit

---

📚 Documentation

Full bilingual documentation is online: uuzjw.github.io/wow-agent

• Guide — intro, quick start, commands
• Architecture — state machine, sub-agents, code indexer
• Tools — file ops, diff/undo, MCP
• Safety — sandbox, egress guard, rollback
• 中文文档

---

🧪 Development & Testing

uv sync                      # install dependencies
uv run python tests_smoke.py # smoke tests (offline)
uv run python tests_e2e.py   # e2e simulation (fake LLM, offline)

---

🤝 Contributing

All kinds of contributions are welcome!

• 🌐 New model/provider adapters — extend PROVIDERS in config.py
• 🎨 UI/UX improvements — ui.py / tui.py (if revived)
• 🔧 New tools — extend TOOLS_SCHEMA + execute in tools.py
• 🐛 Bug fixes / docs — Issues / PRs welcome

Please make the tests pass before submitting:

uv run python tests_smoke.py
uv run python tests_e2e.py

---

License

MIT © uuzjw

Independently developed — unrelated to other open-source projects with the same name. Please report issues at GitHub Issues.