beebots — three AI trading bees, decisions by Jev
An open-source project where three "bees" — Bizzy (breakout, "the grinder"), Breezy (trend, "the calculated one") and Boozy (momentum, "the degen") — trade okx perpetual futures on BTC/ETH/SOL/HYPE. Default mode is paper trading: live prices, simulated money, $333 starting equity per bee. A public dashboard at beebots.tech shows every decision, order, fee and funding payment. The README states plainly what it is: "an experiment and a piece of open-source software," not a trading product, repeated as "not financial advice" three separate times.
Built by a solo developer (GitHub handle imikerussell, first name Mike given only in the README) for the "Creator Magic" YouTube channel, tied to a video titled "I gave three AI bees $1,000." Two days old at ingest (created 2026-09-25) with 103 stars, 57 forks and 9 commits — new, but already the spoke's fastest-growing repo by star velocity.
The decision loop: Jev proposes, code disposes
Every tick, the engine:
- Pulls okx market data — tickers, candles, RSI, MACD, ATR, Bollinger, Donchian, funding rate, open interest.
- Builds a small numeric snapshot from it.
- Asks Jev (TypeSafe AI's typed-decision model — see below) to pick one move, with a probability attached, from a menu of currently valid moves.
- Runs a plain-code risk layer that can veto, shrink or force the move regardless of what Jev picked: max 2x leverage, per-bee stops, a daily loss stop, trade caps, a fee budget, cooldowns, a $1M/24h coin-liquidity gate, and a Jev spend cap ($2/day by default — past it, every bee holds until 00:00 UTC).
- Writes the decision to SQLite before acting, then streams it to the dashboard.
If Jev is slow or unavailable, bees simply hold — Jev is stateless and the README states it "never
sees an order endpoint." Live trading (real money) is opt-in and separately gated: DRY_RUN=false,
MODE=live, three EEA sub-account keys with no withdraw/transfer permission, an explicit
acknowledgment string, and a reduced starting size that ramps up over a configurable window. The
account can never withdraw funds.
Why this matters for agent-deployed-trading: that page's standing worry is "the last human check is an approval prompt" — a human clicks yes once, and a trade has no diff to inspect afterward. Beebots is a different design point: the LLM never chooses freely (Jev answers from a closed menu of valid moves, not free-form tool calls) and the last check is not a human at all but a deterministic code layer running on every single decision, not just the first one. That answers half of agent-deployed-trading's complaint (the guardrail is not "one decision made once, in advance," it fires every tick) and leaves the other half untouched: nothing here proves the risk layer's limits are the right limits, and paper trading being the default sidesteps rather than solves the "a trade has no diff" problem — it just makes the trades free to be wrong.
The Hive: a leaderboard shaped like an overfitting demo
The Hive (beebots.tech, opt-in) is a public leaderboard ranking every user-designed bee by % paper gain or loss, with trades replayed against OKX's public prices to mark a bee "verified." Users design a bee in one sentence; OpenAI turns that into a name, tagline, rule set, allowed-coin list and a portrait modeled on the three official bees' style. Each bee gets one of the three styles above depending on which coins it trades.
Read against this spoke's evaluation corpus, the Hive is a live illustration of backtest-overfitting's mechanism rather than a counterexample to it: an unstated but presumably growing number of independently-parameterized bees run the same three underlying strategies against the same market, and the leaderboard surfaces whichever one got lucky. Nothing published states how many bees exist at a given time (the trial count N that pseudo-mathematics-financial-charlatanism says a Sharpe cannot be assessed without), there's no significance correction, and "verified" here means "replayed against real prices," not "outperformed by more than chance would predict." The Hive is honest about this — its own copy calls it "a game, not a signal service" — which puts it a level above superior-skills's "6 trades, 100% win, validated" framing, but the underlying statistical hazard (best-of-N with N unreported) is the same shape.
What's genuinely different from this spoke's other retail bot platforms
ai-trading-bot-platforms-2026's five platforms (3Commas, Pionex, Cryptohopper, Trade Ideas, SaintQuant) are closed commercial products a buyer cannot inspect. Beebots is MIT-licensed, TypeScript, has tests and CI, and every decision — including the numeric snapshot fed to Jev and Jev's own stated probability — is written to an inspectable SQLite log and streamed live. It sits on trading-bot-platform's control axis close to configurable (a user picks coins and a style sentence, not raw code), but it is free, paper-by-default, and gamified rather than sold as automation — a genuinely new point on that map: open-source, inspectable, LLM-decided, and explicitly a toy before it's a tool.
Stack and dependencies
TypeScript (pnpm, ESLint, tests), a Vite-based dashboard (port 5173), SQLite storage with nightly
7-day backups, Docker Compose deployment with Caddy for HTTPS and GitHub Actions CI. All exchange
access is built on okx's own open-source Agent Trade Kit (MIT): public market data through a
vendored copy of its REST client, signed/order calls through its CLI, targeting the eea.okx.com
endpoint (demo keys supported). Logs redact keys, IPs and emails; a web setup page stays open for two
hours after first start and is claimed by whoever reaches it first, with the owner password stored as
a salted hash.
Related
system-one-models-and-jev · okx · agent-deployed-trading · trading-bot-platform · backtest-overfitting · pseudo-mathematics-financial-charlatanism · ai-trading-bot-platforms-2026 · superior-skills · event-driven-trading · synthesis