Crypto trading bot for your Mac that reads the news before it buys. Practice money first, local AI (Novaeon Sentinel 9B), MetaMask + Hyperliquid.
See the codeA crypto trading bot for your Mac that reads the news before it buys. It trades one simple, tested breakout strategy on Hyperliquid, asks a local AI model (Novaeon Sentinel 9B) whether there is serious bad news about a coin before every purchase, and shows everything in a clear app. It starts with practice money, runs entirely on your Mac, and your wallet key never leaves MetaMask.
By Novaeon Studio. Free and open source (GPL-3.0).
Apple Silicon Mac, macOS 14 or newer. All versions: Releases. Prefer Terminal? See Install.
Risk warning. Trading crypto can lose you money, quickly. Backtests and practice results do not guarantee future results. This is not financial advice. Read the disclaimer before using real money.
![]() | ![]() |
![]() | ![]() |

Screenshots use sample data in practice mode.
The installer checks your Mac's memory and downloads the Sentinel build that fits.
| Your Mac | Sentinel build | Download | News filter on 495 held-out test cases | AI leverage |
|---|---|---|---|---|
| 16 GB or more | 8-bit | 8.4 GB | Same quality as the full model: caught 23 of 28 bad-news cases, blocked 8 of 467 harmless ones | Available (off by default) |
| 8 GB | 4-bit (distilled) | 4.5 GB | Same filter quality as the full model: caught 23 of 28 bad-news cases, blocked 6 of 467 harmless ones | Locked off (its leverage picks are less precise) |
For comparison, the full-precision model catches 24 of 28 and blocks 6 harmless cases. Method and all numbers: docs/SENTINEL.md.
Honest speed note for 8 GB Macs: the 4-bit model uses about 5 GB of memory while idle. In our stress test, with 495 news checks sent back to back, it peaked at about 7.6 GB. The bot works, but the Mac has little room left: close heavy apps and expect other apps to feel slower. A news check only happens right before a purchase (a few times a day), so trading itself is not slowed down. Seconds per check on a base 8 GB M1/M2 have not been measured yet.
Intel Macs are not supported. The installer is Mac-only today: the app and the bot are portable, but the model builds and background services are made for Apple Silicon.
One line (in Terminal):
curl -fsSL https://raw.githubusercontent.com/NovaeonStudio/novaeon-trading-ai/main/install.sh | bash
Prefer to read it first? Download install.sh, read it, then run bash install.sh.
Or the app: download NovaeonTradingAI-<version>.dmg from the
latest release and drag the app to
Applications. The app is not signed with an Apple developer certificate, so macOS blocks the first launch:
If you prefer not to do this, use the one-line install above instead. On first launch it runs the same installer in Terminal; later launches open the app in your browser. Each release lists SHA-256 checksums of its files.
What the installer does:
~/NovaeonTradingAI (its own Python, the trading engine, the app) and starts the
engine, the control service and Sentinel as background services, so trading continues when the browser is closed;~/NovaeonTradingAI/login.txt);Afterwards, ~/NovaeonTradingAI/bin/novaeon controls everything: status, start, stop, logs, password,
update, uninstall. Running the install command again updates an existing install and keeps your settings and
trades. uninstall asks whether to keep your trade history; the bot key in the Keychain is never deleted
automatically.
Real money runs on Hyperliquid (USDC perpetual futures, isolated margin), connected through MetaMask:
REAL MONEY.Start small. Only use money you can afford to lose.
Checked against the code in bot/strategies/. Full rules and research:
docs/STRATEGY.md.
Backtest of these rules (1×, 20 coins, 8 slots, 2,000 USDC start, trading fees and funding included; the news filter and AI leverage cannot be backtested and are not part of these numbers):
| Period | Data | Return | Max drawdown | Trades | Win rate | Profit factor |
|---|---|---|---|---|---|---|
| 2024-11-01 to 2025-09-24 | Binance USDT futures | +64.9% | 16.1% | 663 | 40% | 1.34 |
| 2025-09-24 to 2026-09-25 | Binance USDT futures | +74.5% | 17.5% | 397 | 37% | 2.00 |
| 2025-01-01 to 2026-09-26 | Hyperliquid* | +82.9% | 9.1% | 340 | 38% | 2.27 |
* Hyperliquid only serves about 5,000 hourly candles per coin: 9 of the 20 coins have data for the whole period, the other 11 only from March 2026.
Most trades lose; the strategy lives on a few large moves. Expect long flat or losing stretches (the Hyperliquid test had one that lasted 90 days). Past results do not guarantee future results. Not financial advice.
Sentinel is the news judge. It is a fine-tune of Kev-9B by Jared Palmer, which is built on Qwen3.5-9B-Base. It does not write text: it reads the headlines and answers two typed questions with probabilities. Is there major negative news about this coin? Is the outlook clearly positive, neutral or mixed, or negative?
novaeon-kev/. The raw training dataset is not published.Why fine-tune: on 248 held-out test cases the untuned Kev-9B flagged 26 harmless cases as major bad news and was right about "clearly positive" only 41% of the time. Sentinel flagged 4 harmless cases and was right about "clearly positive" 83% of the time. At the bot's settings, across 495 held-out cases, it catches 24 of 28 bad-news cases and blocks 6 of 467 harmless ones.
Read the full disclaimer.
flowchart LR
subgraph Mac["Your Mac (services listen on 127.0.0.1 only)"]
UI["App<br/>simple and pro mode"]
ENG["Trading engine<br/>+ strategy"]
CTL["Control service<br/>wallet, practice money,<br/>manual buys and stops"]
SEN["Novaeon Sentinel 9B<br/>MLX, 8-bit or 4-bit"]
KC[("macOS Keychain<br/>bot key")]
DB[("SQLite<br/>trades")]
end
MM["MetaMask<br/>(your key)"]
HL["Hyperliquid<br/>prices and orders"]
RSS["News RSS feeds"]
UI -- REST --> ENG
UI -- REST --> CTL
UI -- "sign ApproveAgent" --> MM
ENG -- "before each buy" --> SEN
ENG -- "headlines" --> RSS
ENG --> DB
ENG -- "prices, orders in live mode" --> HL
CTL --> KC
CTL -- "register bot key" --> HL
More detail: docs/ARCHITECTURE.md.
| Path | What it is |
|---|---|
src/, public/, tests/, e2e/ | The app (Vue 3, TypeScript, Nuxt UI, ECharts) and its tests |
bot/strategies/ | The strategy |
bot/bin/ | Control service, start script, watchdog, data tools |
bot/config/ | Configs without keys (dry_run: true) |
novaeon-kev/ | Sentinel training pipeline, labeling guide, evaluation, metrics, model card |
docs/ | Strategy, Sentinel, architecture, security, disclaimer, contributing |
Requirements: Node.js 22 or newer (CI uses 24 and 26) and pnpm (the version is pinned in package.json;
corepack enable picks it up).
pnpm install --frozen-lockfile
pnpm run dev # app with hot reload
pnpm run typecheck
pnpm run lint
pnpm run test:unit
pnpm run build # production build in dist/
pnpm run test:e2e-chromium # browser tests (needs Google Chrome)
The bot side is Python 3.12. Backtests, the Sentinel pipeline and the MLX builds are described in docs/STRATEGY.md and docs/SENTINEL.md. Contributions are welcome: see CONTRIBUTING and the Code of Conduct. Please report security issues privately, as described in SECURITY.
NovaeonTradingAI stands on open-source work, and we are grateful for it:
NovaeonTradingAI is independent and not affiliated with or endorsed by these projects.
Made by Novaeon Studio. Market data and trading via Hyperliquid. News from the public RSS feeds of CoinDesk, Cointelegraph, Decrypt and The Block.
Vue
55.8%
TypeScript
28.7%
Python
11.5%
Shell
3.0%
Crypto trading bot for your Mac that reads the news before it buys. Practice money first, local AI (Novaeon Sentinel 9B), MetaMask + Hyperliquid.
See the codeA crypto trading bot for your Mac that reads the news before it buys. It trades one simple, tested breakout strategy on Hyperliquid, asks a local AI model (Novaeon Sentinel 9B) whether there is serious bad news about a coin before every purchase, and shows everything in a clear app. It starts with practice money, runs entirely on your Mac, and your wallet key never leaves MetaMask.
By Novaeon Studio. Free and open source (GPL-3.0).
Apple Silicon Mac, macOS 14 or newer. All versions: Releases. Prefer Terminal? See Install.
Risk warning. Trading crypto can lose you money, quickly. Backtests and practice results do not guarantee future results. This is not financial advice. Read the disclaimer before using real money.
![]() | ![]() |
![]() | ![]() |

Screenshots use sample data in practice mode.
The installer checks your Mac's memory and downloads the Sentinel build that fits.
| Your Mac | Sentinel build | Download | News filter on 495 held-out test cases | AI leverage |
|---|---|---|---|---|
| 16 GB or more | 8-bit | 8.4 GB | Same quality as the full model: caught 23 of 28 bad-news cases, blocked 8 of 467 harmless ones | Available (off by default) |
| 8 GB | 4-bit (distilled) | 4.5 GB | Same filter quality as the full model: caught 23 of 28 bad-news cases, blocked 6 of 467 harmless ones | Locked off (its leverage picks are less precise) |
For comparison, the full-precision model catches 24 of 28 and blocks 6 harmless cases. Method and all numbers: docs/SENTINEL.md.
Honest speed note for 8 GB Macs: the 4-bit model uses about 5 GB of memory while idle. In our stress test, with 495 news checks sent back to back, it peaked at about 7.6 GB. The bot works, but the Mac has little room left: close heavy apps and expect other apps to feel slower. A news check only happens right before a purchase (a few times a day), so trading itself is not slowed down. Seconds per check on a base 8 GB M1/M2 have not been measured yet.
Intel Macs are not supported. The installer is Mac-only today: the app and the bot are portable, but the model builds and background services are made for Apple Silicon.
One line (in Terminal):
curl -fsSL https://raw.githubusercontent.com/NovaeonStudio/novaeon-trading-ai/main/install.sh | bash
Prefer to read it first? Download install.sh, read it, then run bash install.sh.
Or the app: download NovaeonTradingAI-<version>.dmg from the
latest release and drag the app to
Applications. The app is not signed with an Apple developer certificate, so macOS blocks the first launch:
If you prefer not to do this, use the one-line install above instead. On first launch it runs the same installer in Terminal; later launches open the app in your browser. Each release lists SHA-256 checksums of its files.
What the installer does:
~/NovaeonTradingAI (its own Python, the trading engine, the app) and starts the
engine, the control service and Sentinel as background services, so trading continues when the browser is closed;~/NovaeonTradingAI/login.txt);Afterwards, ~/NovaeonTradingAI/bin/novaeon controls everything: status, start, stop, logs, password,
update, uninstall. Running the install command again updates an existing install and keeps your settings and
trades. uninstall asks whether to keep your trade history; the bot key in the Keychain is never deleted
automatically.
Real money runs on Hyperliquid (USDC perpetual futures, isolated margin), connected through MetaMask:
REAL MONEY.Start small. Only use money you can afford to lose.
Checked against the code in bot/strategies/. Full rules and research:
docs/STRATEGY.md.
Backtest of these rules (1×, 20 coins, 8 slots, 2,000 USDC start, trading fees and funding included; the news filter and AI leverage cannot be backtested and are not part of these numbers):
| Period | Data | Return | Max drawdown | Trades | Win rate | Profit factor |
|---|---|---|---|---|---|---|
| 2024-11-01 to 2025-09-24 | Binance USDT futures | +64.9% | 16.1% | 663 | 40% | 1.34 |
| 2025-09-24 to 2026-09-25 | Binance USDT futures | +74.5% | 17.5% | 397 | 37% | 2.00 |
| 2025-01-01 to 2026-09-26 | Hyperliquid* | +82.9% | 9.1% | 340 | 38% | 2.27 |
* Hyperliquid only serves about 5,000 hourly candles per coin: 9 of the 20 coins have data for the whole period, the other 11 only from March 2026.
Most trades lose; the strategy lives on a few large moves. Expect long flat or losing stretches (the Hyperliquid test had one that lasted 90 days). Past results do not guarantee future results. Not financial advice.
Sentinel is the news judge. It is a fine-tune of Kev-9B by Jared Palmer, which is built on Qwen3.5-9B-Base. It does not write text: it reads the headlines and answers two typed questions with probabilities. Is there major negative news about this coin? Is the outlook clearly positive, neutral or mixed, or negative?
novaeon-kev/. The raw training dataset is not published.Why fine-tune: on 248 held-out test cases the untuned Kev-9B flagged 26 harmless cases as major bad news and was right about "clearly positive" only 41% of the time. Sentinel flagged 4 harmless cases and was right about "clearly positive" 83% of the time. At the bot's settings, across 495 held-out cases, it catches 24 of 28 bad-news cases and blocks 6 of 467 harmless ones.
Read the full disclaimer.
flowchart LR
subgraph Mac["Your Mac (services listen on 127.0.0.1 only)"]
UI["App<br/>simple and pro mode"]
ENG["Trading engine<br/>+ strategy"]
CTL["Control service<br/>wallet, practice money,<br/>manual buys and stops"]
SEN["Novaeon Sentinel 9B<br/>MLX, 8-bit or 4-bit"]
KC[("macOS Keychain<br/>bot key")]
DB[("SQLite<br/>trades")]
end
MM["MetaMask<br/>(your key)"]
HL["Hyperliquid<br/>prices and orders"]
RSS["News RSS feeds"]
UI -- REST --> ENG
UI -- REST --> CTL
UI -- "sign ApproveAgent" --> MM
ENG -- "before each buy" --> SEN
ENG -- "headlines" --> RSS
ENG --> DB
ENG -- "prices, orders in live mode" --> HL
CTL --> KC
CTL -- "register bot key" --> HL
More detail: docs/ARCHITECTURE.md.
| Path | What it is |
|---|---|
src/, public/, tests/, e2e/ | The app (Vue 3, TypeScript, Nuxt UI, ECharts) and its tests |
bot/strategies/ | The strategy |
bot/bin/ | Control service, start script, watchdog, data tools |
bot/config/ | Configs without keys (dry_run: true) |
novaeon-kev/ | Sentinel training pipeline, labeling guide, evaluation, metrics, model card |
docs/ | Strategy, Sentinel, architecture, security, disclaimer, contributing |
Requirements: Node.js 22 or newer (CI uses 24 and 26) and pnpm (the version is pinned in package.json;
corepack enable picks it up).
pnpm install --frozen-lockfile
pnpm run dev # app with hot reload
pnpm run typecheck
pnpm run lint
pnpm run test:unit
pnpm run build # production build in dist/
pnpm run test:e2e-chromium # browser tests (needs Google Chrome)
The bot side is Python 3.12. Backtests, the Sentinel pipeline and the MLX builds are described in docs/STRATEGY.md and docs/SENTINEL.md. Contributions are welcome: see CONTRIBUTING and the Code of Conduct. Please report security issues privately, as described in SECURITY.
NovaeonTradingAI stands on open-source work, and we are grateful for it:
NovaeonTradingAI is independent and not affiliated with or endorsed by these projects.
Made by Novaeon Studio. Market data and trading via Hyperliquid. News from the public RSS feeds of CoinDesk, Cointelegraph, Decrypt and The Block.
Vue
55.8%
TypeScript
28.7%
Python
11.5%
Shell
3.0%