I Built an AI-Native Trading Engine in Python. Here's How It Works

9 strategies, 24/7 WebSocket monitoring, AI scoring, and full agent-based orchestration. Open source, MIT.

Why I Built This

Trading bots are either black-box subscriptions at $50/month, or DIY scripts that break at 3 AM. I wanted a system that:

• Runs unattended — fire and forget
• Understands context — not just "RSI < 30 = buy", but scoring across 9 metrics
• Responds to voice/text — "go AVAX x10" and the AI agent executes everything
• Is open source — so anyone can fork and customize

That's how bybit-ws was born — an AI-native trading engine for Bybit.

Architecture at a Glance

┌─────────────────────────────────────────────────┐
│                  Hermes Agent                     │
│  (AI orchestrator: voice, text, auto-entry)      │
└──────────────┬──────────────────────────────────┘
               │ MCP / REST
┌──────────────▼──────────────────────────────────┐
│              bybit-ws (systemd)                   │
│  main.py → 30s cycle                             │
│  ├── auto_entry.py   (LONG via scoring)           │
│  ├── auto_short.py   (SHORT Tier A/B + junk)      │
│  ├── auto_sl.py      (BB-based stop-loss)         │
│  ├── auto_tp.py      (auto take-profit)           │
│  ├── bb_scalp.py     (⚡ x10 M5 scalping)         │
│  ├── mean_revert.py  (⚡ x10 extreme reversals)    │
│  ├── funding_entry.py(⚡ x10 funding momentum)     │
│  ├── dca.py          (ladder entries)              │
│  ├── rpc.py          (HTTP-RPC on port 8766)      │
│  └── dashboard.py    (SVG dashboard)              │
└──────────────┬──────────────────────────────────┘
               │ WebSocket
┌──────────────▼──────────────────────────────────┐
│               Bybit API v5                        │
└─────────────────────────────────────────────────┘

Core principle: one module = one file. Not a 5,000-line monolith, but 30+ files with clear responsibilities.

9 Strategies — From Conservative to Extreme

The golden rule: the system doesn't guess. Every entry passes through scoring across 8+ metrics:

• Tier bonus (S/A/B/C/D token)
• BB% position (0–100%)
• 24h volume
• RSI(14)
• Funding rate
• BB Squeeze (band compression)
• Consecutive down days
• Weekly/monthly BB%

Minimum threshold: 5.5/10. Anything below — the system silently skips.

AI Integration: More Than Buttons

The interesting part is the AI layer. I use Hermes Agent (by Nous Research) as the orchestrator:

• MCP Server: the agent sees positions, orders, and metrics through a standardized protocol
• REST API: port 8766, CORS, rate-limiting, RPC token
• Auto-entry: the agent decides on entries based on scoring
• GridSignal Bot (@GridSignalBot on Telegram): real-time scanner, LONG/SHORT/x10 buttons

Example interaction:

Me: "go AVAX x10 at market"
Agent: checks BB%, RSI, funding, correlations
       → scoring 7.2/10 → OK
       → calculates: margin=$5, qty=15.1, SL=−5%
       → order placed: 53e99c39

All via voice or text. While driving, on shift, in the shower — doesn't matter.

Numbers & Hardware

• Python 3.11, systemd service
• WebSocket — real-time data from Bybit
• 30-second light cycle — positions, orders, health
• 7-minute heavy cycle — scoring, BB, RSI, squeeze, auto-entry
• Memory: ~200 MB at runtime
• Uptime: weeks without restart

Roadmap

• 🔜 Telegram Mini App — dashboard right in Telegram
• 🔜 Paper trading mode — risk-free testing
• 🔜 Docker one-liner — docker run and done
• 🔮 OKX/Binance support — multi-exchange

Try It

git clone https://github.com/poliakarmai/bybit-ws
cd bybit-ws
cp config.example.yaml config.yaml
# insert your Bybit API keys
systemctl --user enable --now bybit-ws

GitHub: poliakarmai/bybit-ws
Bot: @GridSignalBot
License: MIT

Questions? Open an issue, message the Telegram bot, or fork and customize. The code is alive, trades real money, and improves every day.