FIELD NOTE 003 / OPEN-SOURCE TRADING

25 open-source trading tools we actually read the code of

A source-level review of trading agents, prediction-market infrastructure and quant engines — focused on reusable data pipelines, risk controls and research tooling.

Tensorsoft · 30 October 2026

SOURCE-LEVEL SCREENING48 reviewed → 25 selected

25 selected23 excluded

01 / READ THE CODE, NOT THE PITCH

What survives a closer look?

We cloned 48 repositories and reviewed their source-level implementations, focusing on the code behind “free data” claims. These 25 made the selection; 23 were excluded. The useful parts are often smaller than the product around them: a data adapter, a timestamping rule, a risk veto or a reproducible replay engine.

The notes below reflect the supplied Tensorsoft review. Live endpoint checks were made where possible during that review, not repeated for this publication. “Keyless” does not mean unrestricted, permanently free or production-ready.

02 / THE REPOSITORIES

AI & agent research

01

TradingAgents

TauricResearch

LangGraph firm-sim: analyst → researcher → trader → risk agents. Best feature is the zero-key sentiment trio: Reddit via search.rss (JSON 403s — RSS is the trick), StockTwits public streams, Polymarket Gamma read as market-implied odds. Its settle-against-returns decision log is the cleanest published way to test whether your LLM adds alpha.

02

TradingAgents-CN

hsliuping

The China fork that outgrew upstream: complete map of keyless CN endpoints (AKShare→eastmoney/xueqiu, CLS telegraph) plus the header/referer tricks that keep them alive. Layered vendor-failover with completeness checks before agents see data — reusable for any multi-vendor pipeline.

03

QuantDinger

OpenByteInc

Full-stack AI quant platform whose crown jewel is honesty machinery: walk-forward folds with deflated Sharpe and probability-of-backtest-overfitting — rarely seen outside paid platforms. qlib Alpha158/360 port without the dependency, live Alpaca/IBKR brokers, public funding+OI series behind circuit breakers.

04

OpenAlice

TraderAlice

Local-first orchestrator containing a from-scratch OpenBB clone: typed keyless providers for FRED, BLS, ECB, IMF, OECD, EIA, Deribit chains — exact URLs in code. Vendors self-describe coverage and keyless-state to agents; also ships a working fix for ccxt's broken Hyperliquid markPrice.

05

ai-hedge-fund

virattt

The famous LLM "hedge fund" (Buffett/Munger/Druckenmiller personas). Teaching repo, not trading repo — but two steals: point-in-time fundamentals filtered on filing date (no lookahead) and an event-study module for PEAD-style research. README honestly says so.

06

Vibe-Trading

HKUDS

Multi-market agentic platform (US/CN/A-shares/crypto) whose loader layer alone justifies the clone: no-key Polymarket/Yahoo/Binance/OKX/Eastmoney feeds plus a factor zoo (Alpha101/GTJA191/qlib158) with IC benchmarking. Ignore the agent theater, keep the playbooks.

07

AI-Trader

HKUDS

The arena behind ai4trade.ai — agents trading crypto + prediction markets on a live leaderboard. Two free gifts: drop-in Polymarket/Hyperliquid pricing code (ID resolution, book mid, back-marked candles) and the public leaderboard API, effectively a labeled agent-behavior dataset for studying how bots fail.

08

AutoHedge

The Swarm Corporation

The most marketing-heavy of the bunch — LLM swarm over Jupiter swaps, no risk engine underneath. One honest use: the Jupiter price/ultra client files are a 30-minute Solana pricing starter. Included as a calibrated example of how thin "AI hedge fund" usually is.

09

Hyper-Alpha-Arena

HammerGPT

The closest thing here to a working quant desk: HL + Binance public data (OI, funding, L2, trades) feeding a factor zoo with effectiveness scoring and market-flow triggers. App-shaped (FastAPI+SQLite), so budget extraction time — market_flow_collector is the part worth taking.

10

CloddsBot

alsk1992

One repo with the broadest keyless-endpoint atlas anywhere: Polymarket gamma+CLOB+data-api trio, Kalshi public API with WS, Reddit hot.json, Binance premiumIndex, FRED, CoinGecko, Fear&Greed — each with working request code, not docs. If you clone one repo for data mining, this is it.

11

daily_stock_analysis

ZhuLinsen

Multi-market LLM daily briefing (A/HK/US/JP/KR/TW) on proven-working keyless quote endpoints (sina, tencent, eastmoney screener — exact GET formats verified). The prompt→decision→push structure is a good template for turning noisy tape into one calm morning message.

12

ai-market-maker

olaxbt

LangGraph multi-desk stack with the right instinct: a hard Risk-Guard with veto power over the LLMs. HL public-data adapter, funding/OI CSVs for 20 perps frozen offline, reproducible pinned-data backtest pattern. Paid Nexus dependency limits it; free Binance/HL paths cover most.

13

ContestTrade

FinStep-AI

Multi-agent stock picking run as an internal contest between signal sources — clever mechanism for combining heterogeneous evidence. Practically: two working keyless CN news scrapers (Sina roll, 10jqka push) — exactly the free event streams most pipelines lack. Execution advisory-only.

14

FinMem

pipiku915

Research code behind the FINMEM paper: LLM agent with layered memory scored by exponential decay + importance + access reinforcement. That retrieval math is the reusable core for any agent deciding which past experiences get budget. Backtest is replay only — a citation, not a product.

15

nof0

wquguru

Open clone of the NOF1 LLM arena: Go backend, live Hyperliquid via keyless /info (funding/OI/volume universe), plus NOF1's public leaderboard API as a free meta-signal on how other agents are positioned. Cheapest known way to benchmark prompt strategies against real money.

16

quant-mind

LLMQuant

Not a trading system — a knowledge engine turning papers/news/filings into typed, cited, as-of-timestamped knowledge cards. The timestamps are the point: the right storage shape for news factors without lookahead contamination. Keyless PRNewswire/RSS/arXiv fetchers work today.

03 / THE REPOSITORIES

Prediction markets

17

agents

Polymarket

Polymarket's own AI framework — light demo, but authoritative: gamma.py is the reference for Gamma pagination/filter conventions, and the CLOB sampling endpoints are a free way to enumerate which markets currently carry liquidity rewards. Learn the venue's rules from the venue.

18

prediction-market-analysis

Jon-Becker

+ the data: ships a 36 GiB labeled Polymarket/Kalshi history tarball, free, no crawling — instant calibration/Brier studies. If you self-collect, its RPC log-scanner + keyset cursor is the proven recipe for the complete tape; Kalshi elections endpoints add a second venue for cross-calibration.

19

poly_data

warproxxx

The complete Polymarket fill tape done right: on-chain OrderFilled via Envio HyperSync joined to CLOB metadata, backtrader-ready output. Also documents the CTF Exchange v2 migration date that silently breaks naive historical joins. A self-hostable version of a paid product.

20

poly-maker

warproxxx

Production maker-only engine for Polymarket political markets, built around liquidity-reward economics: two-bid quoting, share merges, scoring bands — the actual math of farming rewards without getting adversely selected. Its batched-metadata trick saves real API budget in any monitor.

21

pmxt

pmxt-dev

Unified prediction-market SDK across eight venues (Polymarket, Kalshi, Limitless, Myriad, Metaculus…) with OpenAPI specs — and the specs are the gold: data-api endpoints for holders, OI, live volume, leaderboards, plus exact WS auth frames nobody had documented.

22

Prediction-Markets-Trading-Bot-Toolkits

HarrierOnChain

Rust/tokio bot framework (copy-trading, whale signals, arb, MM). Ignore the Telegram-marketing wrappers; what survives review is an exact spec of CLOB L2 HMAC signing outside official SDKs and a working Polygon WS log-subscription design for real-time whale tracking.

04 / THE REPOSITORIES

Research & execution engines

23

qlib

Microsoft

Microsoft's quant research platform — the methodology donor of the open quant world: expression engine, factor/cache infra, Alpha158, nested execution. Bar/daily granularity, no tick/L2 replay, crypto path weak. Steal its design, not its data pipeline.

24

ccxt

ccxt

The universal faucet: one API over 100+ exchanges in seven languages — the layer most AI tools above quietly sit on. Not a backtester; acquisition + execution. For anything CEX-perp (bars, funding, OI, books), one client beats forty venue SDKs. MIT.

25

nautilus_trader

Nautech Systems

The most serious engine on the list: Rust core, tick/L2/L3 book replay, identical strategy code in backtest and live, adapters for Binance/OKX/Deribit/Hyperliquid/Tardis — and Polymarket. The door if funding studies ever move from bars to book replay. LGPL: review the repository’s current license and your distribution obligations before incorporating it into closed-source tooling.

05 / METHOD & LIMITATIONS

A reading list, not a performance leaderboard

The supplied review reports cloning and reading the code, with live endpoint checks where possible. The 23 excluded repositories were mostly README-led “AI trading” projects over thin data wrappers, dead backends or referral funnels. Inclusion means there was something useful to study — not that every project is suitable for live trading.

Capabilities, coverage, dataset sizes and comparative judgments are the review author’s observations, not independently verified current guarantees. No commit hashes or endpoint-check timestamps were supplied.
Public endpoints can change, require credentials, impose rate limits or restrict redistribution. Confirm current terms before building on them.
Backtesting methods and agent leaderboards do not establish profitability. Live execution, leverage and market making can cause financial loss.
Inspect code and dependencies before running a repository. Test offline or in a sandbox; do not start with funded wallets or unrestricted trading keys.
Check each repository’s current license and dependency obligations before reuse. The license notes above are not legal advice.

For research and informational purposes only. Not investment advice or an endorsement of any trading system.

Tensorsoft / Less data, higher quality. Back to research