TradingAgents

What it is

TradingAgents runs a LangGraph pipeline of language-model agents over one ticker and one date. Four analysts — market, sentiment, news, fundamentals — write reports; a bull and a bear researcher argue over them; a research manager picks a side; a trader converts the plan into a Buy, Hold or Sell with an entry price, a stop and sizing guidance; three risk debaters argue again; and a portfolio manager closes on a five-tier rating of Buy, Overweight, Hold, Underweight or Sell with an optional price target. Debate and risk rounds are configurable — one each by default, five each on the CLI's "Deep" preset — and every leg is its own model call.

Apache-2.0 by the LICENSE file, roughly 105,000 stars, not archived. Last release v0.4.0 on 31 August 2026; last commit to main 7 September 2026, merging an untagged v0.4.2 branch, with pull requests still being closed on 13 September 2026. 179 issues and 204 pull requests are open.

Pricing

Nothing to buy. Tauric Research publishes no price page at all — /pricing and /plans both 404 on 13 September 2026 — and offers "enterprise access" only through a contact form. The bill is your own model key: gpt-5.6 and gpt-5.6-luna by default, or Anthropic, Google, xAI, DeepSeek, Qwen, GLM, MiniMax, Groq, Mistral, NVIDIA, OpenRouter, Bedrock, Azure, Ollama and any OpenAI-compatible endpoint.

Data & coverage

It ships no data of its own. Prices, indicators, fundamentals and news default to yfinance and need no key; Alpha Vantage is the alternative and does; FRED macro needs a free key; Polymarket, StockTwits and Reddit's public RSS need none. Any market Yahoo covers works through the exchange suffix — 0700.HK, 7203.T, AZN.L, RELIANCE.NS, 600519.SS — with the alpha benchmark chosen per market. Forex and commodity symbols normalise to Yahoo's forms, but the fundamentals leg is equity-shaped and has little to read outside stocks.

Limitations

• The output is a recommendation. The schema carries a Buy/Overweight/Hold/Underweight/Sell rating, an entry price, a stop-loss, sizing such as "5% of portfolio", and a price target. Nobody stands behind any of it; the project's own disclaimer says it is not investment advice.
• No execution. The README's "sent to the simulated exchange and executed" has no code behind it — no broker, no fills, no position accounting anywhere in the repo.
• Not reproducible, by the project's own account: two runs of the same ticker and date can differ, and the curated reasoning models largely ignore the temperature knob.
• Not a backtester either. A historical date is point-in-time guarded — v0.4.0 fixed look-ahead in FRED vintages, social sentiment and the decision memory — but there is no portfolio simulator and no P&L, and the README says backtest results are not guaranteed to match any published figure, the paper's headline return included.
• backtrader is a declared dependency and is imported nowhere in the package.
• pip install tradingagents installs somebody else's package. The PyPI name is held by an unrelated third party at version 0.7.0 under MIT, advertised with the same summary line. The install is a clone plus pip install ., or the Docker image.
• No MCP server, despite the category, and no API to call — a CLI and a Python import.

Alternatives

FinRobot covers similar ground and renders a formal analyst report instead of a debate transcript; FinGPT is the model layer rather than the harness. For market data wired into an LLM without an opinion attached, the servers in the MCP and AI agents category do that job and stop there.