Most trading repos in this roundup do not ask a model to predict anything. Code assembles the state, Jev answers one typed question about it, and code keeps sizing, stops, and keys. What separates the repos is what they feed that decision point and what they refuse to let it touch.
The same question at five venues
| Project | Venue | Jev's answer | Execution |
|---|---|---|---|
| jev-trader | Kuru MON-USDC on Monad | buy or sell per block | post-only order each block |
| Jev Trade | Hyperliquid, five sleeves | long/short, open/close/hold, leverage | dry run without PRIVATE_KEY |
| warrenduffer | Nifty 50 via Kotak/Zerodha | ranker plus take-profit calls | real orders only |
| Polymarket BTC 5m | 5-minute Up/Down rounds | side plus confidence | fill-or-kill via LIVE_TRADING=1 |
| beebots | OKX perpetuals | one move from the legal set | paper default, real prices |
AI Stock, QuantStudio, and Jev-Trades repeat the shape as paper-only ledgers.
One question per tick
jev-trader is the cleanest per-tick decision: one Jev call every Monad block, roughly 300 ms apart, on Kuru's MON-USDC market. It asks about the move over HORIZON_BLOCKS (default 100, about thirty seconds), then posts a post-only limit order. Set MODEL=jev and TYPESAFE_AI_API_KEY; with no PRIVATE_KEY it is a dry run, and the default mock is a momentum heuristic. Dry-run: book read p50 18 ms, loop p50 100 ms.
Jev Trade moves that loop to Hyperliquid as five isolated sleeves (BTC, ETH, SOL, DOGE, BNB), each with its own wallet and Jev call. The answers are Choice options — long/short, open/close/hold, and leverage — and hold is first-class: it sends no order. Entries are post-only Alo quotes one tick inside the touch; exits are Ioc orders that cross it.
AI Hedge Fund takes the opposite route into the same contract: its native JevLLM adapter calls https://api.typesafe.ai/v1/systemone with model jev-1.13.0 by default, so an investor agent runs on Jev's typed decisions like any chat provider. It retries on 429 and 5xx. MIT-licensed and explicitly educational — no trades, no investment advice — it needs a Financial Datasets API key plus one model provider key, with TypeSafe one option among several.
Code keeps the risk
QuantDinger is the clearest gate. Jev sits in front of strategy and Quick Trade entry orders as an auditable PASS/REJECT gate; exits and protective orders bypass it deterministically. The prepared decision context goes to the configured provider through the System One HTTP contract at POST /v1/systemone. Behind it: PostgreSQL state, audit logs, and Prometheus metrics; install is one Docker Compose command, and paper trading is the documented starting point.
warrenduffer gives the model less: every 15 seconds Jev ranks Nifty-50 names through the Vercel AI Gateway (model typesafe-ai/jev via experimental_evaluate, 2,500 ms timeout), then code sizes positions, places stops, and flattens at 15:10 IST. Defaults — capital 100,000, three positions, daily loss cap 1,000 — live outside the model, alongside a kill-switch file and per-position stops. There is no paper mode; only the Kotak path is verified live.
Prediction markets reward abstention
Polymarket BTC 5m Jev trader asks Jev every five seconds whether BTC finishes UP or DOWN versus the window open, and enters only when confidence is above 0.90, ask at most 0.70, edge at least 0.10, and 90+ seconds remain. Anything else and it waits — a confidence gate priced from the cost of being wrong, as confidence gates in production argues. Live orders need LIVE_TRADING=1, with no license file.
The sibling jev-poly-crypto-demo is stricter about settlement: it derives round ids from UTC time, verifies the 300-second market, resolves against the market's own settlement source, and enters only on fee-inclusive edge. It is paper only — live order mode is rejected at startup — and its author warns that class scores are not calibrated win rates and simulated fills ignore queue and competition. Abstention is the default position here; designing for rejection covers that discipline.
Social feeds as optional state
TradingAgents uses Jev as optional preprocessing: with TYPESAFE_API_KEY set, irrelevant StockTwits and Reddit posts drop before the Sentiment Analyst reads them, and stance counts summarize the rest. TYPESAFE_DEFAULT_MODEL can pin jev-1.13.0 instead of the jev-latest alias. Its 108,478 stars are mostly for the agent framework, not Jev.
Jev X Sentiment Analysis turns that into a product: 50 to 1,000 tweets per request alongside CCXT price, funding-rate, and open-interest data, then four typed questions answered concurrently — trade action, sentiment spectrum, squeeze risk, catalyst significance — and a decision card with entry ranges, stops, and targets. It executes nothing; searches cost roughly $0.009 to $0.15 depending on sample size.
What the caveats say
The repos that can touch money ship dry-run defaults; the one-click deploys carry the sharpest warnings. beebots runs on paper out of the box — real OKX prices, $333 per bee, Jev spending capped at two dollars a day — and its opt-in Hive leaderboard shares every paper fill. But setup has no authentication: the page is open to whoever reaches the server first, closes for good once you press Start, and self-locks after two hours, so anyone who wins that race can replace your keys.
Licensing differs in ways worth reading: Jev Trade's README states MIT while GitHub metadata reports NOASSERTION, and the Polymarket trader has no license file at all. None of it prices assets: the gate answers eligibility, the ranker answers order, the screen answers relevance, and the P&L column belongs to code and the market. AI Hedge Fund disclaims trades, advice, and guarantees, and the star counts measure frameworks more than Jev.
What is left is still useful: a decision model earns its place by answering one well-posed question at a fixed cost per evaluation, while everything that can lose money stays in code the repo can test. Copy that seam, and start on the dry run.
