Dart-veider01/TradingAgents-crypto

What it is

This repository is a crypto-oriented fork of TauricResearch's TradingAgents. The upstream framework runs an LLM-driven research desk: analyst agents (technical, news, sentiment, and for equities fundamentals), a bull/bear research debate, a trader agent, and a risk team plus portfolio manager that ends in a Buy, Hold, Sell, Overweight, or Underweight rating. The fork keeps that research layer and adds an execution layer through the bytx library, which can route the final decision to Binance, Bybit, OKX, BingX, KuCoin, Bitget, and other exchanges. Equities are skipped at the execution stage; only crypto tickers are sent.

Quick start

Python 3.11 or later is required. Install the package with the optional execution extra:

pip install ".[bytx]"

Set exchange credentials and enable the driver. Keep demo mode on until you intend to trade real funds:

BYTX_API_KEY=
BYTX_SECRET=
BYTX_PASSWORD=
BYTX_EXCHANGE=bybit
TRADINGAGENTS_EXECUTION_ENABLED=true
TRADINGAGENTS_EXECUTION_DEMO=true
TRADINGAGENTS_EXECUTION_NOTIONAL=100

BYTX_PASSWORD is the API passphrase; only OKX, KuCoin, Bitget, and BloFin require it. Bybit and Binance need only key and secret. Then run a decision that also executes:

tradingagents --ticker BTC-USD --execute

From Python, the same flow uses TradingAgentsGraph and .propagate("BTC-USD", "2026-11-14", asset_type="crypto"), printing the decision and state.get("execution_result"). Paper trading needs no keys: set config["execution"]["paper"] = True. Install the base package without [bytx] if you only want research output.

How orders are placed

A limit within 20 basis points of the market is sent post-only so it does not cross the spread; a farther limit stays a normal limit; with no entry price a market order is used. Default notional is 100 USDT, raised to the venue minimum but never more than three times the requested size. Buy and Overweight open a long; Underweight and Sell close the position; Hold and REVIEW send nothing. Stop and target levels come from the agents.

Requirements and limitations

The full research pipeline needs an LLM provider key (OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Mistral, Groq, NVIDIA, or a local server such as Ollama or an OpenAI-compatible endpoint), or a ChatGPT Plus/Pro subscription login. Crypto runs skip the fundamentals analyst. Live execution is real money and the README explicitly warns it can lose money; the documentation is not financial advice. Results are non-deterministic because LLM sampling varies and news/social data changes over time, though a reasoning model that ignores temperature reduces repeatability. The fork's own contribution is the bytx execution integration; the rest of the documentation mirrors upstream TradingAgents. Licensed Apache-2.0.

HotRepo's take

HotRepo first saw this repository on the 2026-11-15 Python new-project list, with 100 stars and one fork gained in a window of roughly four hours starting that morning. There is no cross-language daily rank on record. With a single short observation window and a brand-new repository, the data supports only that a small audience found it early, not that the project is proven; the differentiating claim, execution through bytx, is documented but not independently verified here.

Project source and documentation