TradingAgents will not tell you what to buy. It will make bull and bear argue first
TradingAgents Report is a hosted multi-agent equity research desk. Dated evidence, a bull/bear debate, risk review, and a five-tier rating. Research only. No live trades.
TradingAgents Report
Updated
The public report library on TradingAgents Report currently opens on Airbnb.
Analysis date: 24 Aug 2026. Exchange: NASDAQ. Portfolio Manager rating: Hold. The decision brief is not a slogan. It is a dated position:
Hold current position, restructured as 70% core long-term stake and 30% tactical trade with defined risk parameters.
Reference price 187.3 USD. Stop or reduce at 184 USD. Conviction: medium. Market and sentiment bullish, news bearish, fundamentals neutral. The page labels the disagreement as a signal conflict, and it states when the thesis dies: a high-volume break below 184.
That is the product in one screen. Not “AI likes Airbnb.” A research snapshot with a date, two sides, and an invalidation line.
Most chat tools will still answer “should I buy this.” The paragraph comes back fluent. It has no analysis date, no listing venue, nobody assigned to argue the other side. The next session the price has moved and the text has not.
TradingAgents Report is a hosted multi-agent research desk. One confirmed listing becomes evidence, debate, risk review, and a five-tier final rating. The homepage is blunt about the boundary: research only. No trade execution.
Open-source engine, hosted desk
The research engine is public at github.com/tradingagents-report/tradingagents-report under Apache License 2.0. Clone it, bring LLM and market-data keys, run the CLI. Redis and PostgreSQL are not required for a local run.
That clone is not this website. TradingAgents Report on this domain is the hosted desk: accounts, usage credits, a job queue, MCP, Simulated Trading, and a curated public report library. The academic paper (arXiv:2412.20138) and the Tauric Research repository remain the origin of the graph. Paper backtests are not product performance.
Reports do not connect to a broker, hold client funds, or place live orders. Simulated Trading is a separate virtual ledger.
Twelve roles, one inspectable path
A run is not one model talking to itself. Specialist modules can collect market, sentiment, news, and fundamentals. Bull and Bear researchers always debate. A Research Manager drafts a plan. A Trader proposes Buy / Hold / Sell as an intermediate step. Aggressive, Conservative, and Neutral risk voices pressure-test it. The Portfolio Manager issues the product rating: Buy, Overweight, Hold, Underweight, or Sell.
The output is a structured decision card first: rating, conviction, reference price, position, zones, target, key risks, invalidation. Full role Markdown sits underneath and can be exported. Unrun modules are marked unavailable instead of being invented.
Public reports without an account
The library is not mock content. Only completed reports the team publishes appear there. On the day these screenshots were taken, that was about 115 reports, filterable by NASDAQ, NYSE, HKEX, SSE, SZSE, TSE, TWSE, and others on the whitelist.
Ratings are not all Hold. The same catalog includes Sells. Each page keeps its analysis date as a data boundary. Public pages do not auto-update when the market moves. A fresh run belongs on the desk.
The research desk
The app is at agent.tradingagentsreport.com. New accounts get analysis credits. Home is a research overview: plan, simulated equity, latest reports, and a start-analysis control.
The desk is built around listing identity, not a free-text ticker box. You confirm the exchange so 700 and 00700 do not collapse into one name, and vendor-only codes stay separate from the public listing.
Evidence modules are optional. Debate, risk review, and the Portfolio Manager always run. A job can cover up to five symbols; each gets its own analysis and credit check. Quick-think and deep-think models come from the admin-open catalog and stay on the same provider for that job.
Jobs queue, run, succeed, fail, or get cancelled. Progress and an event log are visible. Below is AMD at 10%, with the Market Analyst calling indicators. System failures are not charged. A user cancel can still settle actual AI cost.
A stock page keeps quotes, a chart, published reports, and simulated orders on the same listing. Quotes carry source and time. The session in this shot is marked market closed.
Simulated Trading is a ledger, not a brokerage
A five-tier rating is still a hypothesis. Simulated Trading records what you would have done against delayed quotes, with 10,000 USD virtual cash, market and limit orders, and T+0 / T+1 settlement by market. Fees settle in credits, not simulated cash. The server is the account of record.
Reports never place those orders for you. Equity, cash, holdings, and unrealized P&L live on the server.
Credits and coverage
The public catalog is monthly capacity, settled from actual AI cost rather than one point per report. Live checkout is inside the app.
- Starter 20: $20 / month, 2,000 plan points per cycle
- Growth 50: $50 / month, 5,000 plan points per cycle
- Scale 100: $100 / month, 10,000 plan points per cycle
- Max 200: $200 / month, 20,000 plan points per cycle
Signup bonus points can carry over. Plan points refresh each billing cycle. Balance has to clear an analysis threshold before a job starts.
Enabled exchanges currently seed US (NASDAQ, NYSE, AMEX, OTC), Hong Kong, China A-shares, Japan, and Taiwan. Live coverage follows the deployment whitelist. Quotes carry source and time, and can be delayed.
The UI is multilingual. Report language is independent of interface language.
This is a poor fit if you want auto-execution, a broker hookup, or a backtest sold as expected return. It is a better fit if you already ask models about stocks and want the date, the other side, and the invalidation written down.
The product’s own line is enough: research reference, not personalized advice. Check the important numbers yourself.