CyberCode CLI — Codex-dark Web UI with a Built-in Self-Evolving Agent

A standalone, self-contained web UI + agent framework styled after the Codex-dark interface. No external agent dependency — the entire agent core (LLM client, agent loop, 9 atomic tools, layered memory) ships inside agent_core.py and runs from a single directory.

Install

npm install -g cybercode-cli

Then run:

cybercode

That's it. On first run it will:

1. Find Python 3.11+ on your system
2. Install requests if missing
3. Copy the bundled agent + skills to ~/.cybercode/
4. Create a mykey.json template (edit it with your API key, or log in via the web UI)
5. Start the web UI and open your browser

Update

CyberCode automatically checks for updates on startup. To manually update:

cybercode update

Or via npm:

npm update -g cybercode-cli

Diagnostics

Run cybercode doctor to check your environment:

cybercode doctor

Checks Node.js, Python, requests package, working directory, and npm registry version.

Usage

cybercode                                # start the web UI (default)
cybercode webui --port 8080              # custom port
cybercode webui --host 0.0.0.0           # listen on all interfaces
cybercode webui --no-browser             # don't auto-open browser
cybercode webui --dir ~/my-agent         # custom working directory
cybercode webui --llm 1                  # start on 2nd configured LLM
cybercode update                         # self-update to latest version
cybercode doctor                         # run diagnostics

Options

Attribution & License

The agent core architecture in agent_core.py was developed with reference to GenericAgent by lsdefine (MIT License, Copyright © 2025 lsdefine). We gratefully acknowledge the GenericAgent project — its ~100-line agent loop, 9-atomic-tool design, and layered memory concept directly inspired this implementation.

The bundled HyperFrames skills (skills/) originate from HyperFrames by HeyGen (MIT License).

See LICENSE for the full MIT license text, including the GenericAgent copyright notice as required by its license terms.

What it is

cybercodewebui.py boots a self-contained Agent (from agent_core.py), serves cybercodewebui.html at /, and exposes a JSON + SSE API the page talks to. The UI keeps the Codex-dark look — blue radial-gradient desktop, traffic-light window chrome, dark sidebar with threads + skills, centered Let's build empty state, pill LLM switcher, rounded composer with Local/Worktree/Cloud modes — but every control is wired to the built-in agent.

Standalone: no GenericAgent dependency. The agent core (agent_core.py) is a from-scratch reimplementation that ships in this repo. You do not need GenericAgent installed. The only runtime dependency is requests (for the LLM HTTP client); everything else uses the Python stdlib.

Files

cybercode-cli/
├── package.json       # npm package config (bin, files, metadata)
├── bin/
│   └── cli.mjs        # Node.js launcher (finds Python, bootstraps, opens browser)
├── python/
│   ├── agent_core.py       # Self-contained agent core (LLM client + loop + 9 tools + memory)
│   ├── cybercodewebui.py   # stdlib HTTP server + SSE + API
│   └── cybercodewebui.html # Codex-dark UI (single file, no build step)
├── skills/             # HyperFrames video skills (12 files, from HeyGen)
├── templates/
│   └── mykey_template.json  # API key template (auto-copied on first run)
├── LICENSE             # MIT (includes GenericAgent + HyperFrames copyright notices)
└── README.md           # This file

Manual (without npm)

If you prefer to run the Python server directly:

# 1. Configure your LLM (any OpenAI-compatible endpoint)
cat > mykey.json << 'EOF'
{
  "llm1": {
    "apikey": "sk-your-key-here",
    "apibase": "https://api.openai.com",
    "model": "gpt-4o",
    "name": "GPT-4o"
  }
}
EOF

# 2. Install requests (only dependency)
pip install requests

# 3. Run
python python/cybercodewebui.py                 # http://127.0.0.1:18600

Supported LLM providers

Any OpenAI-compatible chat completions endpoint works:

• OpenAI (GPT-4o, o1, o3, etc.)
• DeepSeek
• Kimi / Moonshot
• MiniMax
• Local models via Ollama, LM Studio, vLLM, etc.
• Anthropic Claude via OpenAI-compatible proxies
• l0veyou backend (built-in login support)

What works

The 9 Atomic Tools

The agent core exposes exactly 9 tools (same design philosophy as GenericAgent):

HyperFrames — Built-in Video Generation

The HyperFrames skill pack (12 files) is bundled in skills/. Click the Make video button in the sidebar, or type /video <description>. The agent reads the bundled skills and uses the npx hyperframes CLI to render HTML compositions into .mp4 files. Rendered videos appear in the sidebar gallery and can be played inline.

API Reference

License

MIT License — see LICENSE for full text.

中文文档 | English