Claude Bitcoin Trader: How an AI-Powered Trading Bot Works
Claude Bitcoin Trader: How an AI-Powered Trading Bot Works
Stories about trading bots created using Claude and other AI models are appearing more and more frequently on X. The authors demonstrate automated trades, publish the results, and discuss systems capable of independently analyzing the market and trading via exchange APIs.
One such high-profile project is Claude Bitcoin Trader. Unlike most of the flashy screenshots on social media, it has open-source code, so you can analyze how the bot works from the inside.
We examined the project's code, which is published on GitHub, to understand how the bot makes trading decisions, what role artificial intelligence actually plays in those decisions, and what happens from the moment a quote is received until an order is sent to the exchange.
Project source code: Byte-Ventures / claude-trader — the link is provided at the end of the article.
What's Inside Claude Bitcoin Trader
The overall architecture looks like this:
Exchange → Market Data → Indicators → Signal Score → Filters → AI Review → Risk Control → Order
However, artificial intelligence does not generate the primary trading signal. It is first calculated by standard technical analysis algorithms.
How a bot searches for a trade
After receiving data from the exchange, the system calculates five key metrics:
| Indicator | Weight |
|---|---|
| RSI | 25% |
| MACD | 25% |
| Bollinger Bands | 20% |
| EMA | 15% |
| Volume | 15% |
The RSI measures the strength of recent price movements and helps identify overbought and oversold conditions.
MACD analyzes changes in market momentum using moving averages.
Bollinger Bands show the price's position relative to the current volatility range.
The EMA compares fast and slow exponential moving averages. The system uses the 9-period EMA and the 21-period EMA.
Volume is used as additional confirmation of the price movement.
The results are combined into a single metric— the Signal Score.
Its range is approximately from −100 to +100.
For example:
RSI +18 → MACD +15 → Bollinger +12 → EMA +10 → Volume +7 → Signal Score +62
By default, the trading signal threshold is set to 60.
+60 and above → potential purchase
−60 or lower → potential sale
If the required value is missing, the transaction is not created.
Additional Market Analysis
The signal received is not sufficient to open a position immediately.
The bot also checks higher time frames, including the daily and four-hour charts.
If the direction of the main signal aligns with the larger trend, its score may be increased. Conversely, when trading against the trend, the system may lower the Signal Score.
The next step is to determine the market regime.
The bot analyzes trends and volatility to distinguish between a calm market, a trending market, and extreme conditions.
The complete sequence up to this point:
Data → RSI/MACD/Bollinger Bands/EMA/Volume → Signal Score → Higher Time Frames → Market Mode
AI is not yet required at this stage.
Where Artificial Intelligence Is Emerging
Once a potential deal has been identified, the AI Review system can be activated.
It consists of several AI agents.
Three reviewers independently analyze the deal from different perspectives: one looks for arguments in its favor, another conducts a neutral assessment, and the third looks for reasons to decline the deal.
A separate AI Judge processes them afterward.
Trading Signal → Reviewer 1 + Reviewer 2 + Reviewer 3 → AI Judge → Final Decision
AI may recommend skipping a trade, reducing a position, or postponing it.
However, AI Review is an optional module and can be disabled. The basic trading system is capable of operating without it.
The current version of the project also supports the use of various LLMs. Therefore, despite the name “Claude Bitcoin Trader,” the architecture is no longer limited to Claude alone.
What Happens Before Opening a Position
After the signal is analyzed, the risk management system is activated.
The project includes the following:
- maximum daily loss — 10%;
- maximum hourly loss — 3%;
- position size limit;
- Stop Loss;
- Take Profit;
- Trailing Stop;
- Kill Switch;
- Circuit Breaker;
- Verifying the order before execution.
ATR—a measure of current market volatility—is used for Stop Loss and Take Profit.
By default, the Stop Loss can be set at approximately 1.5 ATR, and the Take Profit at 2 ATR.
Only after passing these checks can the bot submit an order via the exchange's API.
The complete architecture looks like this:
Exchange → Data → Indicators → Signal Score → Trend → Timeframes → AI Review → Risk Management → Coinbase/Kraken → Trade
Paper Trading
By default, the bot operates in Paper Trading mode.
It receives real-time market data and calculates signals, but the trades are executed virtually.
To trade with real funds, you must enable Live Trading separately and grant the system API access to the exchange.
This allows developers to first test the strategy's performance without the risk of losing real money.
Cramer Mode
The project includes an experimental mode called Cramer Mode.
He implements the opposite version of the strategy.
If the main algorithm generates a buy signal, the inverted strategy moves in the opposite direction.
Regular bot: BUY → Cramer Mode: SELL
Regular bot: SELL → Cramer Mode: BUY
This mode is designed to test whether the strategy's input signals have predictive value.
What Role Does AI Actually Play?
The architecture of Claude Bitcoin Trader shows that the core trading logic is not based directly on the language model's decisions.
The base signal is generated using classic technical analysis tools:
RSI + MACD + Bollinger Bands + EMA + Volume → Signal Score
After that, the trend and market condition filters are applied.
AI is integrated at the following level:
Signal Score + market data → AI Review → decision on a potential deal
And immediately before execution, a separate level of risk control is in place:
Solution → Risk Management → Order
Thus, Claude Bitcoin Trader is an algorithmic trading system with an additional AI layer, rather than a language model that independently determines the direction of the market.
Conclusion KLJO
Claude Bitcoin Trader is indeed capable of handling the entire process, from receiving market data to automatically placing an order on Coinbase or Kraken.
However, Claude is not the one generating the main trading signal.
It is generated by a combination of RSI, MACD, Bollinger Bands, EMA, volume, trend analysis, and other algorithmic filters. AI is used as an additional layer of verification for trades that have already been placed.
As a result, the system looks like this:
Algorithms detect a signal → AI evaluates it → risk management verifies it → the bot executes the trade.
The main question remains: how well does this system trade?
The only way to answer this question is to test the strategy's results separately on historical and new data, taking into account commissions, slippage, and maximum drawdown.
Source
Claude Bitcoin Trader — Source Code on GitHub
Byte-Ventures / claude-trader — the project's open-source repository, used to analyze its architecture, trading strategy, AI Review, and risk management system.