[
  {
    "repo": "TauricResearch/TradingAgents",
    "stars": 110279,
    "license": "Apache-2.0",
    "primary_language": "Python",
    "created_at": "2024-12-28T03:31:08Z",
    "pushed_at": "2026-10-03T21:30:10Z",
    "archived": false,
    "one_line": "LangGraph multi-agent framework in which LLM analyst, researcher, trader and risk agents debate to produce a buy/hold/sell decision for one ticker and date.",
    "asset_classes": [
      "US equities",
      "global equities via Yahoo Finance tickers (HK, Tokyo, London, India, Canada, Australia)",
      "China A-shares (.SS/.SZ)",
      "crypto (BTC-USD, ETH-USD)"
    ],
    "llm_providers": [
      "OpenAI",
      "Google/Gemini",
      "Anthropic",
      "xAI/Grok",
      "DeepSeek",
      "Qwen (DashScope intl + CN)",
      "GLM/Zhipu",
      "MiniMax",
      "OpenRouter",
      "Mistral",
      "Kimi/Moonshot",
      "Groq",
      "NVIDIA NIM",
      "Azure OpenAI",
      "AWS Bedrock",
      "Ollama (local)",
      "any OpenAI-compatible endpoint (vLLM, LM Studio, llama.cpp)",
      "TypeSafe Jev (optional social-post screening)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "Yahoo Finance (prices, non-US statements, news search)",
      "SEC EDGAR (US fundamentals, no key)",
      "Alpha Vantage",
      "FRED (macro, optional)",
      "StockTwits",
      "Reddit",
      "Polymarket (data vendor)"
    ],
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": "TradingAgents framework is designed for research purposes. Trading performance may vary based on many factors, including the chosen backbone language models, model temperature, trading periods, the quality of data, and other non-deterministic factors. It is not intended as financial, investment, or trading advice.",
    "architecture": "LangGraph pipeline: parallel analyst team (fundamentals, sentiment, news, technical) -> bull/bear researcher debate -> trader -> risk management team -> portfolio manager, with separate quick-think and deep-think model tiers that can run on different providers.",
    "setup_steps": "About 5 steps: clone, Python 3.11+ venv, pip install, export provider + data API keys (or .env), run the `tradingagents` CLI; Docker Compose alternative; Ollama profile for local models; SEC EDGAR needs a user-agent string.",
    "notable_claims": "None. README states: 'Backtest results are not guaranteed to match any published figure. Returns depend on the model, the temperature, the date range, data quality, and the sampling above. Treat the framework as a research scaffold for studying multi-agent analysis, not as a strategy with a fixed, replicable return.'",
    "evidence": {
      "asset_classes": "README 'Markets and tickers': 'TradingAgents works with any market Yahoo Finance covers, using the exchange-suffixed ticker... China A-shares: Shanghai `.SS`, Shenzhen `.SZ`... Crypto: `BTC-USD`, `ETH-USD`'",
      "llm_providers": "README 'Required APIs' env list (OPENAI_API_KEY ... NVIDIA_API_KEY) and 'The framework supports multiple LLM providers: OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen..., GLM (Zhipu), MiniMax (global + China), OpenRouter, Ollama for local models, and Azure OpenAI for enterprise.'",
      "requires_llm_key": "'Set the API key for your chosen provider'; local exception: 'No key is needed for local servers' (Ollama / openai_compatible).",
      "data_sources": "'US company statements come from SEC EDGAR... Other companies' statements come from Yahoo Finance'; env list 'FRED_API_KEY ... FRED macro data (free, optional)', 'ALPHA_VANTAGE_API_KEY'; 'Sentiment Analyst: Aggregates news headlines, StockTwits, and Reddit chatter'; v0.3.0 news: 'FRED and Polymarket data vendors'.",
      "brokers_execution": "No broker integration documented. 'The `.propagate()` function will return a decision.' The only execution mention is conceptual: 'If approved, the order will be sent to the simulated exchange and executed.'",
      "paper_trading": "Not stated as a mode. README describes decisions, a memory log and backtests; the phrase 'sent to the simulated exchange' appears only in the architecture description.",
      "backtesting": "'`run_backtest` runs the same pipeline over a grid of tickers and dates, writes to a memory log of its own, and scores the decisions whose holding window has since traded.' CLI: `tradingagents backtest NVDA,AAPL --start ... --end ... --every 7`",
      "live_trading_warning": "README blockquote under 'TradingAgents Framework' (quoted verbatim above).",
      "architecture": "'We built TradingAgents with LangGraph'; sections Analyst Team / Researcher Team / Trader Agent / Risk Management and Portfolio Manager; 'The quick model serves the analysts, researchers, debaters and trader; the deep model serves the research and portfolio managers.'",
      "setup_steps": "README 'Installation', 'Docker', 'Required APIs', 'CLI Usage' sections.",
      "notable_claims": "README 'Reproducibility' section."
    }
  },
  {
    "repo": "virattt/ai-hedge-fund",
    "stars": 63905,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2024-11-29T16:30:01Z",
    "pushed_at": "2026-10-02T14:19:04Z",
    "archived": false,
    "one_line": "Educational terminal app in which LLM 'investor agents' run paper-traded or backtested funds on stock tickers using Financial Datasets data; it never places real trades.",
    "asset_classes": [
      "equities (ticker-based funds; prices, fundamentals and earnings from Financial Datasets; market not named explicitly)"
    ],
    "llm_providers": [
      "Anthropic",
      "OpenAI",
      "DeepSeek",
      "Google",
      "xAI",
      "Kimi",
      "TypeSafe (Jev)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "Financial Datasets (financialdatasets.ai)"
    ],
    "brokers_execution": null,
    "paper_trading": true,
    "backtesting": true,
    "live_trading_warning": "This project is for educational and research purposes only. Not intended for real trading or investment. No investment advice or guarantees provided. Creator assumes no liability for financial losses.",
    "architecture": "Investor-persona 'investor agents' are assembled into funds (strategies, capital, cadence, tickers) that are stepped through completed market sessions with an approval step and a hash-chained ledger; one model API key serves the agents.",
    "setup_steps": "3 steps: `pipx install aihf`, run `aihf`, enter a Financial Datasets key and one LLM provider key when prompted (saved to ~/.hedge-fund/.env).",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'A Financial Datasets API key, for prices, fundamentals, and earnings.'; 'Build a new fund from the same list: strategies, capital, cadence, tickers, and it is live.'",
      "llm_providers": "'One model API key for the investor agents. Supported providers: Anthropic, OpenAI, DeepSeek, Google, xAI, Kimi, TypeSafe (Jev).'",
      "requires_llm_key": "'It needs: A Financial Datasets API key... One model API key for the investor agents.'",
      "data_sources": "'A [Financial Datasets](https://financialdatasets.ai) API key, for prices, fundamentals, and earnings.'",
      "brokers_execution": "'Note: the system does not actually make any trades.'",
      "paper_trading": "'Paper trading — your funds, each with a ledger of real market days and fake money... an approval step shows the exact decision about to execute before anything trades'",
      "backtesting": "'Backtesting — replay a fund over history and watch its equity curve draw against its benchmark. Results are saved to `~/.hedge-fund/research/`.'",
      "live_trading_warning": "README 'Disclaimer' section.",
      "architecture": "'One model API key for the investor agents'; 'Every session is appended to a hash-chained ledger in `~/.hedge-fund/paper/<name>/`, so NAV is a track record, not a reset.'",
      "setup_steps": "README 'How to Install' and 'API keys'.",
      "notable_claims": "No performance figures in README."
    }
  },
  {
    "repo": "AI4Finance-Foundation/FinRobot",
    "stars": 8163,
    "license": "Apache-2.0",
    "primary_language": "Python",
    "created_at": "2024-02-27T02:30:30Z",
    "pushed_at": "2026-09-28T11:24:54Z",
    "archived": false,
    "one_line": "Agentic equity-research platform in three generations (AutoGen, OpenAI Agents SDK, PydanticAI desktop app) that produces valuations and research reports rather than executing trades.",
    "asset_classes": [
      "US equities (equity research on tickers such as AAPL, MSFT, NVDA, TSLA; SEC EDGAR filings)"
    ],
    "llm_providers": [
      "DeepSeek (V2)",
      "OpenAI (V0/V1/V2)",
      "Anthropic (V2)",
      "Azure OpenAI (V0)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "FMP",
      "Finnhub",
      "yfinance",
      "SEC EDGAR",
      "Adanos",
      "NewsAggregator",
      "FX",
      "Reddit and FinNLP (V0 data_source/)"
    ],
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": "Disclaimer: The code and documents provided here are released under the Apache-2.0 license. They should not be construed as financial advice or recommendations for live trading. Exercise caution and consult qualified financial professionals before any trading or investment decisions.",
    "architecture": "V2: a Lead Agent orchestrates five role-based pipeline agents (Data -> Analysis -> Modeling -> Synthesis -> Report) plus Bull, Bear and Judge debate agents on PydanticAI; all numbers come from 32 deterministic Python operators and the LLM only narrates.",
    "setup_steps": "V2 local web UI: `uv sync`, `npm install`, `./dev.sh`, then paste an LLM key in Settings (4 steps); desktop .dmg is macOS Apple Silicon only and not notarized (needs `xattr -cr`); V1 needs config.ini with fmp_api_key + openai_api_key; V0 needs conda + OAI_CONFIG_LIST + config_api_keys files.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'finrobot research AAPL # full 13-chapter research artifact'; example reports NVDA, MSFT, TSLA, META, COP; 'Data infrastructure | 7 providers with failover — FMP, Finnhub, yfinance, SEC EDGAR, Adanos, NewsAggregator, FX'",
      "llm_providers": "finrobot_desktop/README.md: 'open Settings and paste your own LLM API key (DeepSeek / OpenAI / Anthropic)'; '--model to pick the LLM (e.g. `anthropic:claude-sonnet-4-6`)'; top README V1: 'edit config.ini: fmp_api_key, openai_api_key'; V0: 'cp OAI_CONFIG_LIST_sample OAI_CONFIG_LIST # OpenAI / Azure OpenAI endpoints'",
      "requires_llm_key": "finrobot_desktop/README.md: 'FinRobot orchestrates your LLM account; it does not ship a key. Until a key is set, the app opens fine but analysis requests return a \"configure your API key\" notice instead of running.'",
      "data_sources": "Top README 'Codebase snapshot' table and repository layout 'data_source/ # finnhub / finnlp / fmp / sec / yfinance / reddit'",
      "brokers_execution": "No broker integration in README; product is report generation: 'turns a ticker into a shareable HTML/PDF report'.",
      "paper_trading": "Not stated.",
      "backtesting": "V0 tutorial 'agent_trade_strategist.ipynb # strategy writing and backtesting'; finrobot_desktop/README.md: 'finrobot backtest ... # needs the `backtest` extra' and notebooks covering 'backtesting'.",
      "live_trading_warning": "README 'License' section disclaimer.",
      "architecture": "README 'Multi-agent architecture' diagram and 'Deterministic compute, LLM narration' section.",
      "setup_steps": "README 'Getting started' sections for V2, V1, V0.",
      "notable_claims": "No return/Sharpe figures in README."
    }
  },
  {
    "repo": "hsliuping/TradingAgents-CN",
    "stars": 32218,
    "license": "NOASSERTION",
    "primary_language": "Python",
    "created_at": "2025-06-26T07:45:09Z",
    "pushed_at": "2026-09-22T12:11:05Z",
    "archived": false,
    "one_line": "Chinese A-share research assistant derived from TradingAgents: multi-agent debate analysis, natural-language stock screening, paper trading and courses on a FastAPI/Vue/MongoDB/Redis stack under a hybrid license.",
    "asset_classes": [
      "China A-shares only"
    ],
    "llm_providers": [
      "OpenAI",
      "DeepSeek",
      "Qwen (通义千问)",
      "others ('等')"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "Tushare",
      "AKShare",
      "BaoStock"
    ],
    "brokers_execution": null,
    "paper_trading": true,
    "backtesting": null,
    "live_trading_warning": "TradingAgents-CN Community Edition is ... For learning and research only — not investment advice. / 本系统仅支持 A 股市场研究，输出为 AI 生成的研究参考，不构成投资建议; 不提供目标价、止损止盈、买卖时机等交易指令",
    "architecture": "Role-playing analyst agents (market index, sector, market, fundamentals, news) plus optimistic/cautious researchers collaborate in debate-style research with five depth levels; v3.0 adds a unified AI assistant with autonomous planning, Skills and memory.",
    "setup_steps": "6 documented steps plus self-installed MongoDB 7+ and Redis 6+ (not bundled), Python 3.11 venv, Node 18+ frontend, .env with DB credentials and JWT_SECRET, then configure LLM keys and data tokens in the UI; documentation is Chinese-language; app/, frontend/, core/ are source-available only.",
    "notable_claims": null,
    "evidence": {
      "license": "LICENSE: Apache-2.0 for tradingagents/, cli/, docs/, tests/, web/; '专有组件（Source Available，源码可见但非开源）' for app/, frontend/, core/ — '禁止：再分发、修改后分发、商业使用'. README: 'hybrid-licensed (Apache-2.0 core engine + proprietary source-available application layer)'.",
      "asset_classes": "'multi-agent stock research assistant for the Chinese A-share market'; '本系统仅支持 A 股市场研究'",
      "llm_providers": "'登录后请先在系统设置中配置大模型 API Key（OpenAI / DeepSeek / 通义千问等）'",
      "requires_llm_key": "Same sentence: LLM API key must be configured in system settings after first login.",
      "data_sources": "'行情与财务数据来自 Tushare / AKShare / BaoStock 等公开接口，需自行注册获取 token。'",
      "brokers_execution": "No broker listed; '不提供目标价、止损止盈、买卖时机等交易指令' (does not provide trading instructions).",
      "paper_trading": "'模拟交易：纸面交易与复盘'; English summary: 'natural-language stock screening, paper trading and study courses'",
      "backtesting": "Not stated.",
      "live_trading_warning": "README English summary and '免责声明' section.",
      "architecture": "'由多个扮演不同角色（大盘/板块/市场/基本面/新闻分析师，乐观/审慎研究员等）的 AI 智能体协作完成研究流程'; '单股研究：多智能体辩论式分析（5 级深度）'",
      "setup_steps": "README '快速开始' / '环境要求' / '启动步骤'.",
      "notable_claims": "None in README."
    }
  },
  {
    "repo": "NoFxAiOS/nofx",
    "stars": 13014,
    "license": "AGPL-3.0",
    "primary_language": "Go",
    "created_at": "2025-10-28T07:17:53Z",
    "pushed_at": "2026-10-04T10:32:54Z",
    "archived": false,
    "one_line": "Self-hosted Go/React terminal where a language model runs a continuous crypto-perpetuals trade loop across nine exchanges under code-enforced risk limits.",
    "asset_classes": [
      "crypto perpetuals on nine exchanges",
      "on Hyperliquid also tokenized US equities, commodities, indices, FX and pre-IPO perps (TSLA, NVDA, GOLD, SPX, EUR, OPENAI)"
    ],
    "llm_providers": [
      "DeepSeek",
      "OpenAI",
      "Claude/Anthropic",
      "Qwen",
      "Gemini",
      "Grok",
      "Kimi",
      "MiniMax",
      "custom endpoints",
      "Claw402 (pay-per-call via x402 USDC, no key)"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "Claw402.ai / Vergex data stack (direction board, direction history, cost-basis and liquidation heatmaps, net flow)",
      "exchange feeds"
    ],
    "brokers_execution": [
      "Binance",
      "Bybit",
      "OKX",
      "Hyperliquid",
      "Bitget",
      "KuCoin",
      "Gate",
      "Aster",
      "Lighter"
    ],
    "paper_trading": null,
    "backtesting": null,
    "live_trading_warning": "Automated trading involves substantial risk. AI-driven strategies are experimental and can lose money. Size positions appropriately, understand each venue, and never trade funds you cannot afford to lose.",
    "architecture": "Each 'trader' is a single-model loop (read market structure -> decide -> execute -> record reasoning) that follows the Vergex direction board; a Go runtime risk engine clamps every order (position limits, leverage caps, exchange-side SL/TP, drawdown auto-close, throttling, safe mode, launch preflight).",
    "setup_steps": "One-line curl installer (or Docker / Railway); then register, fund an AI fee wallet with $1+ USDC on Base, connect Hyperliquid and deposit $12+ USDC, start Autopilot. Exchange registration uses referral links with fee discounts that fund development.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'Crypto perpetuals on all nine exchanges. On Hyperliquid, the same runtime also trades tokenized US equities, commodities, indices, FX, and pre-IPO perps — TSLA, NVDA, GOLD, SPX, EUR, OPENAI — alongside crypto.'",
      "llm_providers": "'Eight providers with your own keys — DeepSeek, OpenAI, Claude, Qwen, Gemini, Grok, Kimi, MiniMax — including custom endpoints and model names. Or no keys at all: Claw402 meters model usage per call in USDC over the x402 protocol.'",
      "requires_llm_key": "Default guided launch uses Claw402: 'put $1+ USDC (Base network) in the AI fee wallet it creates for you'; keys are optional ('with your own keys').",
      "data_sources": "'Decisions come from the Claw402.ai · Vergex data stack: a live direction board for the market universe, per-symbol current direction and direction history, cost-basis and liquidation heatmaps, and real-time market net flow.'",
      "brokers_execution": "README 'Register exchanges' table and architecture diagram 'Exchange Connectivity: Binance · Bybit · OKX · Hyperliquid · Bitget · KuCoin · Gate · Aster · Lighter'",
      "paper_trading": "No paper/testnet mode mentioned; first run: 'connect Hyperliquid and deposit $12+ USDC to trade with.'",
      "backtesting": "Not mentioned in README.",
      "live_trading_warning": "README footer <sub> disclaimer, linking DISCLAIMER.md.",
      "architecture": "'Each trader runs a continuous loop — read market structure, decide, execute, record the reasoning — while a Go runtime clamps every order to hard risk limits the model cannot override.' and 'The model proposes. The runtime disposes.' table.",
      "setup_steps": "README 'First run' and 'Install' sections.",
      "notable_claims": "Leaderboard 'ranked by realized return' but no figures in README."
    }
  },
  {
    "repo": "OpenByteInc/QuantDinger",
    "stars": 12565,
    "license": "Apache-2.0",
    "primary_language": "Python",
    "created_at": "2025-12-28T18:39:26Z",
    "pushed_at": "2026-10-06T21:14:39Z",
    "archived": false,
    "one_line": "Self-hosted Docker AI Trading OS for Python strategies: backtesting, parameter evolution, paper/live execution on crypto exchanges, IBKR and Alpaca, plus an MCP agent gateway.",
    "asset_classes": [
      "crypto (Binance, OKX, Bitget, Bybit, Gate, HTX)",
      "stocks via IBKR and Alpaca",
      "forex (GitHub description)"
    ],
    "llm_providers": [
      "OpenRouter",
      "OpenAI-compatible APIs",
      "Google",
      "DeepSeek",
      "Grok",
      "MiniMax",
      "custom endpoints",
      "TypeSafe Jev (decision gate)"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "yfinance",
      "AkShare",
      "exchange/broker feeds via app/data_sources adapters"
    ],
    "brokers_execution": [
      "Binance",
      "OKX",
      "Bitget",
      "Bybit",
      "Gate",
      "HTX",
      "Interactive Brokers (IBKR)",
      "Alpaca"
    ],
    "paper_trading": true,
    "backtesting": true,
    "live_trading_warning": "QuantDinger can submit real orders when live trading is explicitly enabled. Start with paper trading, use restricted API keys, and review the risk and compliance requirements for your jurisdiction. This project does not provide investment advice. ... Trading, including automated and leveraged trading, can result in the loss of some or all capital.",
    "architecture": "Not a debate-style multi-agent system: deterministic Python strategies (Strategy API V2) and risk checks run in dedicated trading/evaluator workers; an optional AI decision gate (TypeSafe Jev or a configured LLM, fail-open) sits in front of live entry orders; an Agent Gateway and MCP server let external coding agents call approved tools.",
    "setup_steps": "One-line installer pulls a GHCR Docker Compose stack (PostgreSQL, two Redis instances, Kafka, several worker containers); requires setting admin credentials, SECRET_KEY and CREDENTIAL_ENCRYPTION_KEY; optional Prometheus/Grafana overlay; web and mobile clients are separate source-available repos.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'Crypto | Binance, OKX, Bitget, Bybit, Gate, HTX, and adapter extensions.' 'Traditional brokers | IBKR and Alpaca workflows.'; GitHub description: 'paper/live trade across crypto, stocks, and forex'",
      "llm_providers": "'AI providers | OpenRouter, OpenAI-compatible APIs, Google, DeepSeek, Grok, MiniMax, and custom endpoints.'; 'JEV-powered pre-trade decisions' section",
      "requires_llm_key": "Required .env values listed as 'SECRET_KEY, CREDENTIAL_ENCRYPTION_KEY, administrator, PostgreSQL, Redis, and Grafana credentials' (no LLM key); 'If no AI provider is available, the order is allowed and the fail-open result is logged'",
      "data_sources": "Acknowledgements list yfinance and AkShare; repository layout 'data_sources/ Raw market-data source adapters' and 'data_providers/ Aggregated market, macro, news, and sentiment providers'",
      "brokers_execution": "'Strategy and integration surfaces' table; layout 'live_trading/ Normalized crypto exchange adapters', 'alpaca_trading/ Alpaca broker integration', 'ibkr_trading/ Interactive Brokers integration'",
      "paper_trading": "'Agent trading is paper-only by default; live access requires both token and server-side authorization.'; 'Start with paper trading'; 'Signal-only virtual accounts ... never connect to a broker or submit live orders.'",
      "backtesting": "'server-side backtesting and experiment workflows'; 'backtest_engine/ Backtest execution components'; 'Strategy evolution ... walk-forward validation and a final blind holdout'",
      "live_trading_warning": "README top blockquote and 'Legal notice and compliance' section.",
      "architecture": "README 'Architecture' mermaid diagram, 'Backend process roles' table, 'JEV-powered pre-trade decisions' and 'AI agents and MCP' sections.",
      "setup_steps": "README 'Quick start' Option A/B and 'Production deployment'.",
      "notable_claims": "'Results compare parameter robustness; they do not forecast future returns.'"
    }
  },
  {
    "repo": "TraderAlice/OpenAlice",
    "stars": 7249,
    "license": "AGPL-3.0",
    "primary_language": "TypeScript",
    "created_at": "2026-02-18T16:42:55Z",
    "pushed_at": "2026-10-08T08:44:25Z",
    "archived": false,
    "one_line": "Local-first TypeScript workspace connecting external coding agents (Claude Code, Codex, Pi) to market data, Git-backed research, scheduled tasks and human-approved broker execution.",
    "asset_classes": [
      "equities, crypto, commodities, forex and macro (GitHub description)",
      "prediction-market research (Auto Prediction workspace)"
    ],
    "llm_providers": [
      "host agent CLIs: Claude Code",
      "Codex",
      "OpenCode",
      "Pi",
      "Oh My Pi",
      "other supported native agents (bring your own model credential or login)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "embedded provider package (packages/opentypebb): bls, cboe, deribit, eastmoney, ecb, econdb, eia, federal_reserve, fmp, imf, intrinio, multpl, oecd, sec, twse, yfinance",
      "hosted TraderHub reference data",
      "broker/exchange K-lines via Unified Trading Account"
    ],
    "brokers_execution": [
      "Broker Packs: CCXT (crypto exchanges)",
      "Alpaca",
      "Interactive Brokers (IBKR)",
      "LeverUp",
      "Longbridge",
      "built-in Mock simulator"
    ],
    "paper_trading": true,
    "backtesting": null,
    "live_trading_warning": "Trading execution is beta. Start with simulator, paper, demo, or testnet accounts. OpenAlice is experimental software; interfaces may change, and it provides no guarantees of correctness, reliability, profitability, or loss prevention. Use real funds only if you understand and accept the risks.",
    "architecture": "An orchestrator ('Alice') hosts external native coding agents inside Git-backed Workspaces (Chat, AutoQuant, Auto Prediction) with scheduled Issues and an Inbox; a Unified Trading Account lets agents stage proposed operations that a human approves via 'Trading as Git'.",
    "setup_steps": "Download the desktop app (macOS/Windows) or `pnpm install && pnpm dev`; connect a model credential or login for at least one host agent CLI; broker SDKs are installed later as separate versioned Broker Packs after choosing a broker; no broker account needed for research.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "README: 'Auto Prediction for prediction-market research'; GitHub description: 'An AI trading agent covering equities, crypto, commodities, forex, and macro'",
      "llm_providers": "'Use Claude Code, Codex, OpenCode, Pi, Oh My Pi, and other supported native agents with Alice's market data'; 'The desktop app includes Pi; connect a model credential or supported login to start researching.'",
      "requires_llm_key": "'Source installs need at least one host agent CLI and its model credentials or login.'",
      "data_sources": "Repo tree: packages/opentypebb/src/providers/{bls,cboe,deribit,eastmoney,ecb,econdb,eia,federal_reserve,fmp,imf,intrinio,multpl,oecd,sec,twse,yfinance}/index.ts; docs/market-data-architecture.md ('TraderHub/reference data', 'Bar service', 'vendor source or UTA broker source'); README: 'Available data depends on your configured providers.'",
      "brokers_execution": "docs/broker-packs.md: 'It does not contain the live broker SDKs for CCXT, Alpaca, IBKR, LeverUp, or Longbridge. Those implementations ship as versioned, platform-specific Broker Packs'; 'Mock is the sole built-in engine.'; packages/uta-broker-{alpaca,ccxt,ibkr,leverup,longbridge}; tests/integration/broker-{alpaca,bybit,hyperliquid,ibkr,okx}-paper",
      "paper_trading": "README caution: 'Start with simulator, paper, demo, or testnet accounts.'; services/uta/src/domain/trading/brokers/mock/MockBroker.ts; ui/src/pages/SimulatorPage.tsx",
      "backtesting": "Not stated. README mentions 'AutoQuant for quantitative projects and experiments'; no path in the repo tree contains 'backtest'.",
      "live_trading_warning": "README '[!CAUTION]' block under 'Bring research to a trading decision'.",
      "architecture": "README 'Features' section: Workspaces, Issues, Inbox, Workspace Manager, Unified Trading Account, 'Trading as Git'.",
      "setup_steps": "README 'Get Started' and 'Development'; docs/broker-packs.md installation transaction.",
      "notable_claims": "None in README."
    }
  },
  {
    "repo": "Lumiwealth/lumibot",
    "stars": 2110,
    "license": "GPL-3.0",
    "primary_language": "Python",
    "created_at": "2020-09-10T10:00:16Z",
    "pushed_at": "2026-10-08T21:58:51Z",
    "archived": false,
    "one_line": "Python trading framework (since 2020) with twelve broker integrations and backtesting, plus a built-in AI agent runtime letting LLM agents research, debate and submit orders.",
    "asset_classes": [
      "stocks",
      "options",
      "crypto",
      "futures",
      "forex",
      "indexes",
      "prediction contracts (Polymarket)"
    ],
    "llm_providers": [
      "OpenAI (default `openai/gpt-6-luna`)",
      "any LiteLLM/ADK-supported provider string",
      "Gemini (used in a recorded run)"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "Yahoo Finance",
      "Alpaca",
      "Interactive Brokers REST",
      "ThetaData",
      "Polygon/Massive",
      "DataBento",
      "Tradier",
      "Schwab",
      "Polymarket price history",
      "CCXT (Kraken, Binance, KuCoin, BitMEX, Bybit, OKX)",
      "Pandas/CSV",
      "agent tools: SEC fundamentals/filings, FRED macro, Alpaca news"
    ],
    "brokers_execution": [
      "Alpaca",
      "Interactive Brokers (and IB REST)",
      "Tradier",
      "Schwab",
      "Tradovate",
      "TopstepX via ProjectX",
      "Bitunix",
      "Polymarket",
      "selected CCXT paths (Coinbase, Kraken, WEEX auto-detected; KuCoin, Binance, BitMEX manual)"
    ],
    "paper_trading": true,
    "backtesting": true,
    "live_trading_warning": "Not research-only. README says: 'Backtests are not expected future performance.' and 'Start with paper trading. When you are ready for live trading, use the same strategy class and switch your broker account/configuration intentionally.'",
    "architecture": "Agents are created inside a normal `Strategy` subclass with `self.agents.create(...)` (e.g. researcher, bull, bear read-only; trader with `allow_trading=True`); a deterministic Python gate and the standard broker/backtest loop place orders, and decisions are recorded to SQLite/Parquet artifacts.",
    "setup_steps": "Two commands for the demo (`pip install lumibot`, `lumibot demo`: no key, no broker); AI backtest needs Python 3.10+, OPENAI_API_KEY and `BACKTESTING_DATA_SOURCE=yahoo`; paper/live needs broker keys; README repeatedly promotes the BotSpot hosted cloud and a ThetaData promo code.",
    "notable_claims": "'Recorded run (made with Gemini, before GPT-6 Luna became the default): ten fresh agent runs across April 6 to 10, with one verified fill for 15 SPY shares.' No return, Sharpe or win-rate figures; 'A fresh model run can choose different trades and returns.'",
    "evidence": {
      "asset_classes": "'Lumibot supports stocks, options, crypto, futures, forex, indexes, and prediction contracts across several broker integrations'",
      "llm_providers": "'The default model is `openai/gpt-6-luna` on medium reasoning; set `model=` per agent to use any other LiteLLM/ADK-supported provider string'",
      "requires_llm_key": "'That runs a real backtest on free daily data and writes a tearsheet. No API key, no broker account, no configuration.'; '`--template python` ... No AI model or model API key is required.'; AI example: 'You need Python 3.10+ and an OpenAI API key (`OPENAI_API_KEY`).'",
      "data_sources": "README 'Select Backtesting Data Sources' list; 'Built-in AI agent tools include ... Alpaca news when credentials exist, technical indicators, SEC fundamentals and filings, FRED macro data'",
      "brokers_execution": "README 'Supported Brokers' list.",
      "paper_trading": "'lumibot run my-bot --paper'; 'ALPACA_IS_PAPER=true'; 'Start with paper trading.'",
      "backtesting": "'`Strategy.backtest(...)` always backtests'; 'Backtest before connecting a broker.'",
      "live_trading_warning": "Quotes from 'Example: Research, Bull, Bear, and Trader Agents' and 'Run the same strategy with a paper broker'.",
      "architecture": "Code example `self.agents.create(name=\"trader\", ..., allow_trading=True)`; 'A deterministic Python gate the model cannot talk past, real broker orders, and a trace you can open.'; 'Memory and Traceability' section.",
      "setup_steps": "README 'Sixty seconds', 'Run your first AI backtest', 'Recommended Data Provider' (promo code BotSpot10).",
      "notable_claims": "README 'Run your first AI backtest' recorded-run paragraph."
    }
  },
  {
    "repo": "microsoft/RD-Agent",
    "stars": 14867,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2024-04-03T03:39:33Z",
    "pushed_at": "2026-10-06T22:45:59Z",
    "archived": false,
    "one_line": "Microsoft's general LLM research-and-development automation agent whose finance scenario, RD-Agent(Q), iteratively proposes and implements Qlib quant factors and models; it does not execute trades.",
    "asset_classes": [
      "equities via Qlib quant factor/model loops ('real stock markets'; specific market not named in README)"
    ],
    "llm_providers": [
      "LiteLLM backend (default) — any LiteLLM-supported model",
      "OpenAI",
      "Azure OpenAI",
      "DeepSeek",
      "SiliconFlow (embeddings via litellm_proxy)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "Qlib data (vendor not named in README)",
      "user-supplied financial reports folder for factor extraction"
    ],
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": null,
    "live_trading_warning": "The RD-agent is aimed to facilitate research and development process in the financial industry and not ready-to-use for any financial investment or advice. ... The RD-agent does not provide financial opinions or reflect the opinions of Microsoft, nor is it designed to replace the role of qualified financial professionals.",
    "architecture": "Two roles in an evolving loop: 'R' proposes hypotheses/ideas and 'D' implements them in runnable code with feedback; finance scenarios (`rdagent fin_quant`, `fin_factor`, `fin_model`, `fin_factor_report`) co-optimize Qlib factors and models.",
    "setup_steps": "Linux only; Docker required and runnable without sudo; conda env (Python 3.10/3.11); `pip install rdagent`; .env with CHAT_MODEL/EMBEDDING_MODEL via LiteLLM (needs ChatCompletion, json_mode and embedding ability); `rdagent health_check`; Streamlit or token-protected Flask web UI for traces.",
    "notable_claims": "'Extensive experiments in real stock markets show that, at a cost under $10, RD-Agent(Q) achieves approximately 2× higher ARR than benchmark factor libraries while using over 70% fewer factors. It also surpasses state-of-the-art deep time-series models under smaller resource budgets.' Backtest period and benchmark are not stated in the README (refers to the NeurIPS 2025 paper arXiv:2505.15155).",
    "evidence": {
      "asset_classes": "'Extensive experiments in real stock markets show...'; 'Run the Automated Quantitative Trading & Iterative Factors Model Joint Evolution: [Qlib] self-loop factor & model proposal and implementation application'",
      "llm_providers": "'We now fully support LiteLLM as our default backend for integration with multiple LLM providers.'; configuration examples for OpenAI, Azure OpenAI, DeepSeek, SiliconFlow embedding.",
      "requires_llm_key": "'The demos requires following ability: ChatCompletion, json_mode, embedding query'; .env example 'OPENAI_API_KEY=<replace_with_your_openai_api_key>'",
      "data_sources": "'rdagent fin_factor_report --report-folder=<Your financial reports folder path>'; Qlib scenarios documented at rdagent.readthedocs.io (data vendor not in README).",
      "brokers_execution": "No broker or execution integration mentioned.",
      "paper_trading": "Not stated.",
      "backtesting": "The word 'backtest' does not appear in the README; evaluation happens inside Qlib factor/model loops.",
      "live_trading_warning": "README '⚖️ Legal disclaimer' section.",
      "architecture": "'we have identified a framework with two key components: R for proposing new ideas and D for implementing them.'",
      "setup_steps": "README '⚡ Quick start' ('RD-Agent currently only supports Linux', Docker installation, Configuration).",
      "notable_claims": "README '🥇 The First Data-Centric Quant Multi-Agent Framework!' section."
    }
  },
  {
    "repo": "pipiku915/FinMem-LLM-StockTrading",
    "stars": 962,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2023-12-01T04:01:38Z",
    "pushed_at": "2024-08-18T01:58:13Z",
    "archived": false,
    "one_line": "Research code for the FinMem paper: a single LLM trading agent with layered memory and character profile, simulated in train/test mode over historical single-stock data.",
    "asset_classes": [
      "US single stocks (TSLA default example)"
    ],
    "llm_providers": [
      "OpenAI (gpt-4 backbone; text-embedding-ada-002 always required)",
      "HuggingFace models served via TGI endpoint (HF_TOKEN)"
    ],
    "requires_llm_key": true,
    "data_sources": null,
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": null,
    "architecture": "One agent with layered memory and a character design; `train` mode populates memory from historical information, `test` mode makes decisions from memory plus new information; checkpoint resume for unstable API runs.",
    "setup_steps": "Docker build from `.devcontainer/Dockerfile` (Python 3.10), .env with OPENAI_API_KEY (+ HF_TOKEN for gated models), edit `config/config.toml` for the backbone endpoint, run `run.py sim`; last push August 2024.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'--start-time ... The training or test start time [default: 2022-06-30 For Ticker 'TSLA']'; config/tsla_config.toml",
      "llm_providers": "'The model can be run with LLMs on HuggingFace that can be deployed via TGI... we will always use the `text-embedding-ada-002` as our embedding model'; 'To run the OpenAI model... model = \"gpt-4\"'",
      "requires_llm_key": "'the `OPENAI_API_KEY` variable needs to be set in `.env` no matter what backbone LLM is used.'",
      "data_sources": "Not stated; input is a preprocessed pickle '--market-data-path ... [default: data/06_input/subset_symbols.pkl]'",
      "brokers_execution": "None mentioned.",
      "paper_trading": "Not stated.",
      "backtesting": "'Usage: run.py sim [OPTIONS] Start Simulation' with --start-time/--end-time over historical dates; 'Run mode: train or test'",
      "live_trading_warning": "No disclaimer in README.",
      "architecture": "'In the train mode, the information populate the agent's memory. In the test mode, the agent will use the information in the memory and new information to make decisions.'; figures memory_flow.png, workflow.png, character.png",
      "setup_steps": "README 'Usage' and 'Build Docker Image & Run the Container'.",
      "notable_claims": "README defers to the paper (arXiv:2311.13743); no figures stated."
    }
  },
  {
    "repo": "alsk1992/CloddsBot",
    "stars": 2924,
    "license": "MIT",
    "primary_language": "TypeScript",
    "created_at": "2026-01-26T15:08:42Z",
    "pushed_at": "2026-10-02T19:37:29Z",
    "archived": false,
    "one_line": "Claude-powered trading terminal (Node.js) for prediction markets, crypto perps and DeFi across 21 chat channels, with 118+ strategies, backtesting, token launches and its own token.",
    "asset_classes": [
      "prediction markets (Polymarket, Kalshi, Betfair, Smarkets, Drift, Opinion.xyz, Predict.fun; data-only Manifold, Metaculus, PredictIt)",
      "crypto perpetual futures (Binance, Bybit, Hyperliquid, MEXC, Drift, Percolator, Lighter)",
      "Solana and EVM DEX spot (Jupiter, Raydium, Orca, Uniswap V3, 1inch...)",
      "Bittensor subnet mining"
    ],
    "llm_providers": [
      "Claude/Anthropic (primary)",
      "GPT-4/OpenAI",
      "Gemini",
      "Groq",
      "Together",
      "Fireworks",
      "AWS Bedrock",
      "Ollama"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "platform orderbooks/candles/price feeds",
      "Chainlink feeds",
      "GoPlus token audits",
      "external: FedWatch, 538, Silver Bulletin, RCP, Odds API"
    ],
    "brokers_execution": [
      "Polymarket",
      "Kalshi",
      "Betfair",
      "Smarkets",
      "Drift",
      "Opinion.xyz",
      "Predict.fun",
      "Binance",
      "Bybit",
      "Hyperliquid",
      "MEXC",
      "Percolator",
      "Lighter",
      "Solana DEXs (Jupiter, Raydium, Orca, Meteora, Pump.fun, Bags.fm)",
      "EVM DEXs (Uniswap V3, 1inch, PancakeSwap on ETH/ARB/OP/Base/Polygon)"
    ],
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": null,
    "architecture": "Four agents (Main on Claude, Trading, Research, Alerts) with 121 skills and 18 tools behind a gateway serving 21 messaging channels; a unified strategy/risk layer (118+ strategies, VaR/CVaR, circuit breaker, Kelly sizing) routes to prediction-market, DeFi and perp exchange adapters.",
    "setup_steps": "Node.js 22+ (20 unsupported); `npm install -g <GitHub release tgz>` then `clodds onboard` wizard; needs ANTHROPIC_API_KEY plus per-venue keys/private keys (e.g. POLYMARKET_API_KEY, SOLANA_PRIVATE_KEY); README also advertises a Clodds token contract address, paid token launches and a compute API.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "README tables 'Prediction Markets (10)', 'Perpetual Futures (7 Exchanges)', 'Crypto & DeFi', 'Bittensor Mining'",
      "llm_providers": "'8 LLM providers: Claude (primary), GPT-4, Gemini, Groq, Together, Fireworks, AWS Bedrock, Ollama'",
      "requires_llm_key": "Configuration: '# Required ANTHROPIC_API_KEY=sk-ant-...'",
      "data_sources": "'Market Data | Real-time orderbooks, candles, liquidity tracking, depth analysis, price feeds across all platforms'; 'External Data: FedWatch, 538, Silver Bulletin, RCP, Odds API for edge detection'; architecture diagram 'Chainlink Feeds'",
      "brokers_execution": "'Trading | Order execution on 16+ platforms (prediction markets, futures, Solana DEXs, EVM DEXs)' and the platform tables.",
      "paper_trading": "Only for arbitrage: 'Note: Defaults to dry-run mode.' No general paper-trading mode stated.",
      "backtesting": "'Backtesting | Configurable strategy backtesting with historical data, SL/TP validation, P&L analysis'",
      "live_trading_warning": "No research-only or financial-risk disclaimer found in README (grep for disclaim/educational/financial advice returned nothing).",
      "architecture": "'AI System: 8 LLM providers, 4 agents: Main, Trading, Research, Alerts, 18 tools' and the 'Architecture' diagram.",
      "setup_steps": "README 'Quick Start', 'Configuration'; header 'Clodds CA: 2puc76eh...pump'.",
      "notable_claims": "No performance figures in README."
    }
  },
  {
    "repo": "hummingbot/condor",
    "stars": 222,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2023-12-28T18:44:45Z",
    "pushed_at": "2026-10-08T14:43:18Z",
    "archived": false,
    "one_line": "Hummingbot Foundation Telegram/web bot for monitoring and trading crypto CEX/DEX venues via the Hummingbot API, with an optional LLM assistant and dry-run-capable domain agents.",
    "asset_classes": [
      "crypto spot and perpetuals on centralized exchanges (Binance, Bybit, etc.)",
      "DEX swaps and CLMM liquidity positions via Gateway"
    ],
    "llm_providers": [
      "OpenAI",
      "OpenRouter (tool-calling models)",
      "custom OpenAI-compatible endpoints (Venice AI, Together, Fireworks, vLLM, LM Studio)",
      "Codex via ACP bridge"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "Hummingbot API exchange connectors",
      "Gateway (DEX)"
    ],
    "brokers_execution": [
      "centralized exchanges via Hummingbot API connectors (Binance, Bybit, etc.)",
      "DEX protocols via Hummingbot Gateway"
    ],
    "paper_trading": true,
    "backtesting": null,
    "live_trading_warning": null,
    "architecture": "Telegram -> Condor Bot -> Hummingbot API -> trading bots (and Gateway -> DEX); optional `/agent` LLM assistant plus domain agents that 'author their own tick strategy (analysis order, decision rules, risk limits)' and run it on a loop.",
    "setup_steps": "curl installer from hummingbot/deploy; Mac/Linux only (WSL2 on Windows); Docker required for the Hummingbot API stack; Telegram bot token or Local mode; Tailscale strongly recommended for production; `make doctor` health check.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'CLOB Trading - Place orders on centralized exchanges (Binance, Bybit, etc.)'; 'DEX Trading - Swap tokens and manage CLMM liquidity positions via Gateway'; '/trade | CEX and DEX trading menu (spot & perpetual orders, positions, swaps)'",
      "llm_providers": "'AI Assistant ... (optional OpenAI or OpenRouter keys, or any custom OpenAI-compatible endpoint like Venice AI...)'; 'Connect Venice AI, Together, Fireworks, or your own vLLM / LM Studio server'; 'Refresh Codex: ... Codex ACP bridge'",
      "requires_llm_key": "'OPENAI_API_KEY=sk-... # Optional, for AI features'",
      "data_sources": "No third-party data vendor named; 'All commands use direct API calls via `hummingbot_api_client`'",
      "brokers_execution": "'Condor controls real trading through Hummingbot API: orders, balances, bots, and stored exchange keys.'; architecture diagram.",
      "paper_trading": "'AI Agents - Domain agents that can be asked, delegated to, or run on a loop — each can author its own tick strategy ... and run it autonomously with dry-run support'",
      "backtesting": "Not mentioned.",
      "live_trading_warning": "No research/education-only disclaimer; README has a security warning: 'Condor controls real trading through Hummingbot API... the risk surface has grown.'",
      "architecture": "README 'Architecture' diagram and 'Features' bullets.",
      "setup_steps": "README 'What you need', 'Install Condor (start here)', 'Secure Connection via Tailscale'.",
      "notable_claims": "None."
    }
  },
  {
    "repo": "mnemox-ai/tradememory-protocol",
    "stars": 1425,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2026-02-23T10:35:46Z",
    "pushed_at": "2026-10-05T08:43:51Z",
    "archived": false,
    "one_line": "Local-first MCP memory server and tamper-evident audit trail for AI trading agents; syncs fills, recalls losing trades before orders and (preview) brakes Alpaca orders.",
    "asset_classes": [
      "US equities (Alpaca; AAPL, TSLA)",
      "crypto (Hyperliquid fills, Binance spot)",
      "forex/gold via MetaTrader 5 (XAUUSD)"
    ],
    "llm_providers": [
      "MCP-client agnostic (Claude Desktop, Claude Code, Cursor)",
      "Anthropic (optional, evolution tools)",
      "DeepSeek (default for replay) or Anthropic"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "Hyperliquid public fills API (no key)",
      "Alpaca (read-only sync)",
      "MT5 and Binance spot sync scripts",
      "Binance public market data (api.binance.com) for evolution tools"
    ],
    "brokers_execution": [
      "Alpaca via the `proxy` extra (preview) that forwards policy-allowed orders to Alpaca's MCP server",
      "MetaTrader 5 (sync/read)"
    ],
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": "For educational and research purposes only. Not financial advice.",
    "architecture": "Not a trading agent itself: a 20-tool MCP server (memory, state, planning, risk, audit, evolution) with five memory layers (episodic, semantic, procedural, affective, trade records) and a Recall -> Record -> Reflect -> Audit loop; the optional brake uses Mnemox Control (AGPL-3.0) as a policy proxy between agent and broker.",
    "setup_steps": "`pip install tradememory-protocol`; add to Claude Desktop / Claude Code MCP config (`claude mcp add tradememory -- uvx tradememory-protocol`); `tradememory sync hyperliquid --address` (no key) or `sync alpaca --env-file`; brake needs Python 3.12+, `[proxy]` extra, Alpaca account id and `proxy doctor`.",
    "notable_claims": "No return claims. 'Empirical validation: ongoing (n=14 trades, target n>=100 for statistical significance; at n=14 the confidence intervals are too wide to conclude anything - see validation/final_verdict.md)'. Brake test: 'run against a real Alpaca paper account on 2026-10-01: three refusals ..., one allowed one-share bracket order that reached the broker'.",
    "evidence": {
      "asset_classes": "'Three ways it is used' table: 'US equity trader | Forex EA system (XAUUSD) | Audit trail'; 'tradememory sync hyperliquid'; 'MT5 and Binance spot sync scripts are in `scripts/`'",
      "llm_providers": "'Add to Claude Desktop'; 'evolution calls the Anthropic API only if you set `ANTHROPIC_API_KEY`. Replay calls DeepSeek by default (or Anthropic), and only with that provider's key set.'",
      "requires_llm_key": "'Memory server (default). Never places orders and never asks for broker keys. Records and recalls only, in a local SQLite file.'",
      "data_sources": "'Hyperliquid: public fills, no key needed'; 'Alpaca: read-only calls with your own keys'; 'The evolution tools read public Binance market data (`api.binance.com`)'",
      "brokers_execution": "'The `proxy` extra runs TradeMemory between your agent and your broker's MCP server (Alpaca today).'",
      "paper_trading": "No own paper mode; brake was tested 'against a real Alpaca paper account on 2026-10-01' (broker-side paper).",
      "backtesting": "MCP tools table: 'Evolution | evolution_fetch_market_data · evolution_discover_patterns · evolution_run_backtest · evolution_evolve_strategy'; 'Evolution engine: research phase'",
      "live_trading_warning": "README footer: 'For educational and research purposes only. Not financial advice.'",
      "architecture": "'How it works' (Recall/Record/Reflect/Audit), 'MCP Tools' table, 'Brake (`proxy` extra)' and license note on Mnemox Control.",
      "setup_steps": "README 'Start with your own history', 'Connect your agent', 'Put a brake in front of your broker (preview)'.",
      "notable_claims": "README 'Research Status' and brake status paragraph."
    }
  },
  {
    "repo": "ryanfrigo/kalshi-ai-trading-bot",
    "stars": 614,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2025-07-26T18:41:32Z",
    "pushed_at": "2026-10-08T17:45:42Z",
    "archived": false,
    "one_line": "Kalshi prediction-market toolkit: signed API client, risk governor, OpenRouter LLM strategy, paper trader, MCP server and an edge harness scoring settled trades against the book.",
    "asset_classes": [
      "US prediction markets (Kalshi event contracts)"
    ],
    "llm_providers": [
      "any model via OpenRouter (default `anthropic/claude-sonnet-4.5`)",
      "Claude skill (.claude/skills/kalshi-trade)"
    ],
    "requires_llm_key": false,
    "data_sources": [
      "Kalshi Events API / settlements (REST + WebSocket)",
      "AAA Fuel Gauge",
      "Ornn OCPI index",
      "OpenRouter rankings chart",
      "Roll Call counts (data-fed market families)"
    ],
    "brokers_execution": [
      "Kalshi (authenticated REST + WebSocket client with RSA signing)"
    ],
    "paper_trading": true,
    "backtesting": false,
    "live_trading_warning": "Read this before running with real money. Nothing here is guaranteed to make money, and this README makes no such promise. The whole point of the toolkit is to let you measure whether you have edge — not to hand you one.",
    "architecture": "One LLM call per decision (README: 'It is not a \"5-model ensemble\" despite earlier README claims'), fractional-Kelly sizing with category/sector guardrails, a risk governor authoritative for every live order, and an adversarial-verify gate (research -> skeptic -> deterministic edge recompute); multi-agent debate scaffolding is quarantined in experimental/.",
    "setup_steps": "Clone, `python setup_env.py` (Python 3.12+), `.env` with KALSHI_API_KEY and OPENROUTER_API_KEY, Kalshi private-key file in project root, `python cli.py health`, then `cli.py run --paper`.",
    "notable_claims": "Negative claim, with sample stated: 'NO MEASURED EDGE — won 58% vs. 60% implied by the book (edge −1.9 pts). Brier 0.2356.' (operator's live account, n=55 forward-settled trades, published 2026-10-02; benchmark = Kalshi book-implied probability, forward-only). Also: Beast Mode — 'Running this with real money historically led to significant losses on this repo.'",
    "evidence": {
      "asset_classes": "'An agent-native toolkit for Kalshi'; 'A Kalshi account with API access'",
      "llm_providers": "'a pluggable LLM client (any model on OpenRouter)'; settings 'primary_model = \"anthropic/claude-sonnet-4.5\"'",
      "requires_llm_key": "'OpenRouter key (only for AI directional strategy)'; 'Safe Compounder ... Pure edge-based math, no LLM required.'",
      "data_sources": "'Market ingestion | Pulls the full tradeable universe via the Events API'; 'Data-Fed Families' table (AAA Fuel Gauge, Ornn OCPI, OpenRouter rankings, Roll Call)",
      "brokers_execution": "'Kalshi client | Authenticated REST + WebSocket client (RSA signing, retries, rate-limit handling)'; 'python cli.py run --live # live trading (real money)'",
      "paper_trading": "'python cli.py run --paper # paper trading'; '## Paper Trading: Simulate trades without sending real orders.'",
      "backtesting": "'Not a backtest. A backtest needs a captured price/outcome corpus this repo doesn't ship yet (`cli backtest` explains why).'",
      "live_trading_warning": "README blockquote after the intro.",
      "architecture": "'1. AI Directional ... it calls a single LLM via OpenRouter (with a fallback chain on errors)'; 'verify | ... a SKEPTIC tries to refute the fade → a deterministic gate recomputes the edge'; 'experimental/ quarantine: UNWIRED scaffolding'",
      "setup_steps": "README 'Quick Start' and 'Installation'.",
      "notable_claims": "README 'Prove Your Edge (the headline)' and 'Example Strategies' sections."
    }
  },
  {
    "repo": "AI4Finance-Foundation/FinRL",
    "stars": 16594,
    "license": "MIT",
    "primary_language": "Jupyter Notebook",
    "created_at": "2020-07-26T13:18:16Z",
    "pushed_at": "2026-10-08T15:07:24Z",
    "archived": false,
    "one_line": "Original deep-reinforcement-learning stock-trading framework (train-test-trade pipeline with Stable-Baselines3 agents), now positioned as educational with production work redirected to FinRL-X/FinRL-Trading; no LLM agents.",
    "asset_classes": [
      "US stocks (DOW 30 tutorial)",
      "CN securities / A-shares (Akshare, Baostock, Tushare, JoinQuant, RiceQuant)",
      "crypto (Binance, CCXT)",
      "FX (FXMacroData)",
      "Taiwan securities (Sinopac)"
    ],
    "llm_providers": [],
    "requires_llm_key": false,
    "data_sources": [
      "Akshare",
      "Alpaca",
      "Baostock",
      "Binance",
      "CCXT",
      "EODhistoricaldata",
      "FXMacroData",
      "IEXCloud",
      "JoinQuant",
      "QuantConnect",
      "RiceQuant",
      "Sinopac",
      "Tushare",
      "WRDS",
      "YahooFinance"
    ],
    "brokers_execution": [
      "Alpaca ('Basic Alpaca support' per the FinRL vs FinRL-X table; trade.py stage)"
    ],
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": "Disclaimer: We are sharing codes for academic purposes under the MIT license. Nothing herein constitutes financial advice or a recommendation to trade real money. Users are solely responsible for any financial decisions made using this software.",
    "architecture": "Three-layer DRL stack — Gym-style market environments, DRL agents (ElegantRL, RLlib, Stable-Baselines3: A2C, DDPG, PPO, SAC, TD3) and financial applications — in a train-test-trade pipeline; no LLM roles.",
    "setup_steps": "Clone, venv, `pip install -e .`, run three example scripts (download DOW 30 data from Yahoo, train five DRL agents, backtest vs MVO and DJIA); README repeatedly redirects new users to FinRL-X.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'This script downloads DOW 30 stock data from Yahoo Finance'; 'Supported Data Sources' table (CN Securities, Cryptocurrency, FX spot rates, Taiwan securities); applications folder 'cryptocurrency_trading', 'portfolio_allocation', 'high_frequency_trading'",
      "llm_providers": "No LLM provider named; 'Strategies | DRL agents (A2C, DDPG, PPO, SAC, TD3)'; FinRL-X described as 'AI-Native (ML + DRL + LLM-ready)'",
      "requires_llm_key": "No LLM key in setup; data download from Yahoo Finance.",
      "data_sources": "README 'Supported Data Sources' table.",
      "brokers_execution": "'Live Trading | Basic Alpaca support | Full multi-account integration + risk controls' (FinRL vs FinRL-X table); unit_tests/downloaders/test_alpaca_downloader.py",
      "paper_trading": "Not stated.",
      "backtesting": "'python examples/FinRL_StockTrading_2026_3_Backtest.py ... compares their performance against two baselines: Mean Variance Optimization (MVO) and the DJIA index.'",
      "live_trading_warning": "README 'LICENSE' section disclaimer.",
      "architecture": "'FinRL is the original open-source framework for financial reinforcement learning, organized around three core layers: Market Environments, DRL Agents, Financial Applications'",
      "setup_steps": "README 'FinRL Stock Trading 2026 Tutorial'.",
      "notable_claims": "No figures in README."
    }
  },
  {
    "repo": "freqtrade/freqtrade",
    "stars": 55143,
    "license": "GPL-3.0",
    "primary_language": "Python",
    "created_at": "2017-05-17T23:48:53Z",
    "pushed_at": "2026-10-08T19:43:01Z",
    "archived": false,
    "one_line": "Mature open-source crypto trading bot (since 2017) with backtesting, hyperopt, dry-run, Telegram/web-UI control and FreqAI adaptive machine learning; no LLM agents.",
    "asset_classes": [
      "crypto spot and futures (Binance, BingX, Bitget, Bybit, Gate, HTX, Hyperliquid, Kraken, OKX; others via CCXT)"
    ],
    "llm_providers": [],
    "requires_llm_key": false,
    "data_sources": [
      "exchange OHLCV via CCXT (`download-data`)"
    ],
    "brokers_execution": [
      "Binance",
      "BingX",
      "Bitget",
      "Bybit (and Bybit EU)",
      "Gate (and Gate EU)",
      "HTX",
      "Hyperliquid (DEX)",
      "Kraken",
      "OKX / MyOKX",
      "community-tested: Bitvavo, Kucoin",
      "futures: Binance, Bitget, Bybit, Gate, Hyperliquid, Kraken, OKX"
    ],
    "paper_trading": true,
    "backtesting": true,
    "live_trading_warning": "This software is for educational purposes only. Do not risk money which you are afraid to lose. USE THE SOFTWARE AT YOUR OWN RISK. THE AUTHORS AND ALL AFFILIATES ASSUME NO RESPONSIBILITY FOR YOUR TRADING RESULTS. Always start by running a trading bot in Dry-Run and do not engage money before you understand how it works.",
    "architecture": "Rule-based Python strategy classes executed by a bot loop with persistence in SQLite; optional FreqAI self-training ML models; control via Telegram RPC or FreqUI; no agent roles.",
    "setup_steps": "Docker quickstart recommended; Python 3.12+; config and strategy files; README: 'We strongly recommend you to have coding and Python knowledge.'",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'Freqtrade is a free and open source crypto trading bot'; 'Supported Spot Exchanges' and 'Supported Futures Exchanges' lists.",
      "llm_providers": "No LLM mentioned; 'Adaptive prediction modeling: Build a smart strategy with FreqAI that self-trains to the market via adaptive machine learning methods.'",
      "requires_llm_key": "No LLM key in README.",
      "data_sources": "'download-data Download backtesting data.'; 'optimize your buy/sell strategy parameters with real exchange data'",
      "brokers_execution": "README 'Supported Exchange marketplaces'.",
      "paper_trading": "'Dry-run: Run the bot without paying money.'",
      "backtesting": "'Backtesting: Run a simulation of your buy/sell strategy.'; CLI `backtesting`, `hyperopt`",
      "live_trading_warning": "README 'Disclaimer' section.",
      "architecture": "README 'Features' and 'Bot commands'.",
      "setup_steps": "README 'Quick start' and 'Disclaimer'.",
      "notable_claims": "None."
    }
  },
  {
    "repo": "ProjectDXAI/continuous-record-llm-trading-agents",
    "stars": 80,
    "license": "CC-BY-4.0",
    "primary_language": null,
    "created_at": "2026-09-01T14:28:45Z",
    "pushed_at": "2026-09-10T13:55:41Z",
    "archived": false,
    "one_line": "Paper-artifact repository (PDF, figures, aggregate data, share cards) for a DXRG study of two production LLM trading-agent fleets; it contains no runnable agent code.",
    "asset_classes": [
      "Base memecoin pools (real ETH) — DX Terminal Pro",
      "Hyperliquid perpetuals — DXAP fleet"
    ],
    "llm_providers": [
      "Qwen3-235B via SGLang (Terminal Pro, frozen runtime)",
      "qwen3.7-plus via OpenRouter (DXAP, 'mostly')"
    ],
    "requires_llm_key": null,
    "data_sources": null,
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": null,
    "live_trading_warning": "The DXAP record in this paper ends on August 15, 2026 and covers the pre-alpha research fleet. It is not a performance summary of the current DXAP Alpha. ... At the fleet level, the pre-alpha record did not show a directional edge.",
    "architecture": "Describes but does not ship the DXAP agents: multi-tool turns under a frozen single-model runtime, a user-set risk slider and a nine-row gainers leaderboard rendered in the prompt; the paper's thesis is that this operating layer drives behaviour more than strategy text.",
    "setup_steps": "Not installable software; read paper.pdf, figures/, data/ extracts, or load the two Hugging Face datasets.",
    "notable_claims": "'43.2% of positions touched +300 bps of open profit within 24h; 49.3% of those still closed negative; median capture was 2.0%. A fixed 2%/4% bracket beat every discretionary exit we measured, +39.0 bps per position.'; 'Median leverage is 5.0x in every volatility sextile'; 'At the fleet level, the pre-alpha record did not show a directional edge.' Periods stated: Feb 26–Mar 18 2026 (3,505 vaults) and Jun 8–Aug 15 2026 (500–599 agents, 231,638 turns, 14,596 fills); no external benchmark.",
    "evidence": {
      "asset_classes": "'3,505 user-funded DX Terminal Pro vaults trading real ETH on Base for 21 days, then a DXAP live alpha fleet of 500 to 599 user-created agents on Hyperliquid perpetuals for 69 days.'",
      "llm_providers": "'The record at a glance' table: 'Model | Qwen3-235B via SGLang, one frozen runtime | Mostly qwen3.7-plus via OpenRouter'",
      "requires_llm_key": "Not applicable — repository holds paper artifacts only.",
      "data_sources": "Not applicable; 'data/ holds the aggregate extracts behind Figures 1, 2, 3, 5, and 8'",
      "brokers_execution": "No code; the studied fleets traded on Base pools and Hyperliquid perps.",
      "paper_trading": "Fleet described as 'Hyperliquid perpetuals, paper at live prices plus a small real-capital book' (not a feature of this repo).",
      "backtesting": "Not applicable.",
      "live_trading_warning": "README 'The paper is the pre-alpha baseline' and finding 4.",
      "architecture": "Finding 1: 'A risk slider explains leverage (+0.425 per level), agent fixed effects absorb 60% of variance, and a nine-row leaderboard rendered in the prompt routes 46.5% of all trade entries.'",
      "setup_steps": "'This repository holds the paper, figures, aggregate extracts, and reusable share cards.'",
      "notable_claims": "README 'Four findings' and 'The record at a glance'."
    }
  },
  {
    "repo": "Harvard-AI-and-Robotics-Lab/TradingAgent",
    "stars": 1,
    "license": "Apache-2.0",
    "primary_language": "Python",
    "created_at": "2026-10-05T01:08:35Z",
    "pushed_at": "2026-10-05T01:09:42Z",
    "archived": false,
    "one_line": "TrustTrade paper code: a TradingAgents fork adding multi-agent credibility scoring, selective consensus, deterministic temporal signals and reflective memory, evaluated on AAPL, GOOG, NVDA.",
    "asset_classes": [
      "US equities (AAPL, GOOG, NVDA)"
    ],
    "llm_providers": [
      "OpenAI (required in quick start, e.g. gpt-4o-mini)",
      "Anthropic (optional)",
      "Google (optional)",
      "xAI (optional)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "Finnhub (FINNHUB_API_KEY)",
      "market, news and social-media data via dataflows/ (provider-restricted caches excluded from release)"
    ],
    "brokers_execution": null,
    "paper_trading": null,
    "backtesting": true,
    "live_trading_warning": "This software is for research and educational purposes only. It is not financial, investment, legal or trading advice. Users are responsible for provider costs, data rights, model outputs and any decisions made with this software.",
    "architecture": "Deterministic temporal signals -> parallel domain analysts -> credibility scorer (selective consensus) -> bull/bear researchers -> trader (BUY/HOLD/SELL + %) -> short/long reflective memory; the upstream risk-debate stage is kept but omitted from the main experiments.",
    "setup_steps": "Clone, venv, `pip install -e .`, `bash scripts/validate_release.sh` (offline), export OPENAI_API_KEY + FINNHUB_API_KEY, `python main.py --ticker NVDA --trade-date 2024-05-10 --mode trusttrade --provider openai --model gpt-4o-mini`; exact paper reproduction marked 🚧 (needs author-approved frozen manifests). Repo created 2026-10-05 with 1 star.",
    "notable_claims": "'market-only signals at the Analyst stage reach 33.1% cumulative return at 9.1% maximum drawdown'; 'Full-allocation agents achieve roughly 30% average cumulative return, but incur approximately 12% maximum drawdown'; 'The high-consensus configuration raises average cumulative return from roughly 10% to 26%, while moving maximum drawdown from about 3% to 8%'. Period stated: 2024 Q1 on AAPL/GOOG/NVDA plus a forward-time window 2026-01-01 to 2026-02-18; reference: rule-based baselines and 19 human annotators. README adds: 'These values summarize the paper experiments; they are not a promise of live trading performance.'",
    "evidence": {
      "asset_classes": "'The paper evaluates AAPL, GOOG and NVDA, across 2024 Q1 historical backtesting and a 2026 Q1 forward-time study'",
      "llm_providers": "'export OPENAI_API_KEY=... export FINNHUB_API_KEY=... # Optional providers export ANTHROPIC_API_KEY=... export GOOGLE_API_KEY=... export XAI_API_KEY=...'",
      "requires_llm_key": "'TrustTrade can make multiple paid model and data-provider calls per trading decision.'",
      "data_sources": "'export FINNHUB_API_KEY=...'; 'dataflows/ online/offline data interfaces'; 'provider-restricted data caches' excluded",
      "brokers_execution": "None mentioned.",
      "paper_trading": "Not stated.",
      "backtesting": "'baseline_analysis/ rule-based and LLM backtesting code'; 'TrustTrade adds ... portfolio-aware backtesting'",
      "live_trading_warning": "README '⚠️ Research-use disclaimer'.",
      "architecture": "README '🧭 Framework' pipeline diagram and '🧠 Agent implementation' table.",
      "setup_steps": "README '🚀 Quick start', '🔐 Configure providers', '🧪 Reproducibility status'.",
      "notable_claims": "README '🔑 Main findings' and '📊 Main results'."
    }
  },
  {
    "repo": "0xethanq/astra-quant-agent",
    "stars": 368,
    "license": "NOASSERTION",
    "primary_language": "Python",
    "created_at": "2026-08-30T14:34:10Z",
    "pushed_at": "2026-10-03T20:02:58Z",
    "archived": false,
    "one_line": "Autonomous crypto-perpetuals system where a multi-LLM council proposes order intents and deterministic Python risk gates plus OKX-native OCO orders execute them every 15 minutes.",
    "asset_classes": [
      "crypto derivatives / perpetuals on OKX"
    ],
    "llm_providers": [
      "DeepSeek-V3/R1",
      "Claude 3.5 Sonnet (Anthropic)",
      "GPT-4o (OpenAI)",
      "Qwen"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "OKX exchange feeds (orderbook, CVD, funding, open interest, IV skew, VWAP factors); no separate data vendor named"
    ],
    "brokers_execution": [
      "OKX only"
    ],
    "paper_trading": true,
    "backtesting": null,
    "live_trading_warning": "This project is an open-source algorithmic trading framework and quantitative research system, created strictly for academic research, education, and simulation purposes. ... Users must maintain comprehensive risk management practices and validate strategies thoroughly in paper trading (`demo`) before committing live capital.",
    "architecture": "Four decoupled boundaries: market ingestion/factors -> multi-model council (Macro, Trend, Microstructure, Contrarian seats with Paranoid Veto / Weighted Majority / Alpha Momentum consensus) -> Python physical risk gates (sizing clamps, R:R floor, circuit breaker) -> OKX execution with cloud OCO brackets.",
    "setup_steps": "Clone, `./setup.sh` wizard, Docker Compose (recommended) or bare-metal Python 3.11+/Node 18+ frontend build; OKX API key with Read+Trade only; admin plane on :8080; README includes exchange referral links; license is AGPL-3.0 + Commons Clause + anti-scam addendum (no resale, no paid signals).",
    "notable_claims": null,
    "evidence": {
      "license": "LICENSE: 'distributed under the GNU Affero General Public License version 3 (AGPL-3.0), subject to the Commons Clause License Condition v1.0 and the Anti-Scam / Financial Fraud Special Addendum'",
      "asset_classes": "'Cryptocurrency derivatives trading entails significant capital risk'; 'OKX-native conditional order execution'",
      "llm_providers": "'Specialized Seats: Configurable models (DeepSeek-V3/R1, Claude 3.5 Sonnet, GPT-4o, Qwen) evaluate distinct domains'",
      "requires_llm_key": "Core loop: '2. Multi-Model Council (LLMs) ──> Macro, Trend, and Microstructure seats analyze & synthesize order intent'; admin plane: 'Risk settings, prompt studio, API keys'; 'LLM Gateway & Thinking Config'",
      "data_sources": "'Market Microstructure Factors' table (CVD, OBI, funding rate, OI, IV skew, VWAP); '443 (TCP Outbound): OKX V5 REST API & LLM gateway connectivity'",
      "brokers_execution": "'AstraQuant's autonomous execution engine only supports OKX native API. Binance and Gate.io are community partner links provided solely for registration discounts'",
      "paper_trading": "'Fail-Closed Default: AstraQuant boots in paper trading mode (`demo`) by default. Live execution is physically blocked until exchange credentials are provided and validated'",
      "backtesting": "Not documented as a feature; a `tests/backtest` suite is referenced and the license text mentions 'individual backtesting'.",
      "live_trading_warning": "README 'Disclaimer' section.",
      "architecture": "README 'System Architecture' and 'Multi-Model Analysis & Prompt Caching'.",
      "setup_steps": "README 'Quick Start', 'Deploy', 'Exchange Referral & Fee Rebates', 'License'.",
      "notable_claims": "No performance figures in README."
    }
  },
  {
    "repo": "yebof/quant-agent",
    "stars": 11,
    "license": "MIT",
    "primary_language": "Python",
    "created_at": "2026-04-12T06:38:05Z",
    "pushed_at": "2026-09-23T18:38:43Z",
    "archived": false,
    "one_line": "Nine-agent LLM swing-trading system for 97 US equities/ETFs: six scheduled daily sessions on Alpaca, schema-enforced reasoning, hard Python risk engine, quarterly self-evolution.",
    "asset_classes": [
      "US equities and ETFs (index, sector and inverse ETFs plus ~80 single names)"
    ],
    "llm_providers": [
      "Anthropic (claude-opus-4-7; stated default since 2026-05-11)",
      "OpenAI (gpt-5.5; stated as 'current default' in Configure section)",
      "DeepSeek (OpenAI-compatible, optional)"
    ],
    "requires_llm_key": true,
    "data_sources": [
      "yfinance",
      "FRED API",
      "RSS feeds (Reuters, CNBC, MarketWatch, AP, BBC, NPR, Fed)",
      "SEC EDGAR",
      "Alpaca API"
    ],
    "brokers_execution": [
      "Alpaca (OTO bracket orders, calendar, live quotes)"
    ],
    "paper_trading": true,
    "backtesting": null,
    "live_trading_warning": "This software is provided for educational and research purposes only. It is NOT investment advice. Trading securities involves substantial risk of loss; you can lose more than your initial deposit. ... By using this software, you accept full responsibility for any trades it places, including against real-money accounts.",
    "architecture": "Eight daily agents (Tech Analyst, News Intelligence, Macro, Earnings, Portfolio Manager, Risk Manager, Position Reviewer, Evening Analyst) plus a quarterly Meta Reflector; parallel analyst fan-out -> PM with an 8-layer memory stack -> hard risk engine -> LLM risk manager with veto -> SELLs-then-BUYs execution, driven by systemd/launchd timers across six ET sessions.",
    "setup_steps": "Python 3.11+, Alpaca + FRED + Anthropic (and/or OpenAI) keys in `.env`, edit `config/settings.yaml`, enable seven systemd timers (or launchd on macOS, which on Sequoia needs Full Disk Access for /bin/bash and plugged-in power); 874 tests. Repo has 11 stars.",
    "notable_claims": null,
    "evidence": {
      "asset_classes": "'LLM multi-agent quantitative trading system for US equities'; 'Trading Universe: 97 symbols ... Index ETFs: SPY, QQQ, IWM, DIA ... Inverse ETFs: SH, SDS, PSQ, SQQQ'",
      "llm_providers": "Prerequisites: 'Anthropic API key (default since 2026-05-11 — all 9 agents on `claude-opus-4-7`) and/or OpenAI API key'; Configure: 'Current default is `gpt-5.5` (OpenAI) for all 9 agents'; .env 'DEEPSEEK_API_KEY=... # optional'",
      "requires_llm_key": "'ANTHROPIC_API_KEY=sk-ant-... # also the cross-provider failover key (OpenAI/DeepSeek → Claude)'",
      "data_sources": "README 'Data Sources' table: yfinance, FRED API, RSS feeds, SEC EDGAR, Alpaca API",
      "brokers_execution": "'execution/broker.py # Alpaca (OTO brackets, calendar, live prices)'; 'BUY attaches OTO stop-loss via Alpaca (broker-enforced)'",
      "paper_trading": "'Default configuration points at Alpaca paper trading; switching to a live account is your decision, and the consequences are yours alone'",
      "backtesting": "Not mentioned in README.",
      "live_trading_warning": "README top '⚠️ Disclaimer' blockquote.",
      "architecture": "README 'Architecture' session diagram and 'Agents' table.",
      "setup_steps": "README 'Setup', 'Production deployment', 'Tests'.",
      "notable_claims": "No return figures; only '874 tests pin every invariant'."
    }
  }
]