How the degen bot works
Two jobs, kept strictly apart: AI picks the token, a script makes every trade. No model ever signs a transaction. Here is the whole machine, top to bottom.
1 · The scout finds a token (every 30 minutes)
- Discover. Pulls a few hundred Solana tokens from DexScreener searches and boosts, Jupiter's top-trending, top-traded and top-organic lists, and GeckoTerminal's trending pools. Free public data, no wallet involved.
- Hard script gates. Liquidity, 24h volume, market-cap band, pair age, two-sided buy/sell flow, not a stablecoin, not a parabolic pump or crash. The micro-cap bag adds rug gates from Jupiter's token audit: mint and freeze authority revoked, top holders ≤ 40%, organic volume.
- Backtest. Three days of 5-minute candles. A calibrator picks the token's market phase and fits a range config; it must hold more tokens than buy-and-hold at real slippage and at double slippage. A mean-reversion (Ornstein–Uhlenbeck) fit measures how fast the price snaps back. Most tokens fail here.
- Cheap AI screen. Claude Haiku scores the survivors 0–10 for fit to this strategy.
- Escalate. Anything scoring high goes to Claude Sonnet for a full analysis: flow quality, liquidity vs trade size, rug and manipulation risk, whether the edge will last. Verdict: ADOPT, WATCH or REJECT.
- Final call. Only when switching the bag to a new token is on the table, Claude Opus decides — switching costs a round trip of slippage, so it must be clearly better.
Token names and descriptions on the web are untrusted text. The AI only ever chooses among tokens that already passed the script gates, and every number the bot trades with comes from the script's backtest — never from the AI. If the AI is unavailable, the scout falls back to script-only picks that are paper-traded only.
2 · The script trades the range (every 60 seconds)
- Range window. It tracks a rolling window of recent prices. Position 0% = the bottom of that range, 100% = the top.
- Buy low. In the bottom zone, after the price ticks up off the low (it never catches a falling knife), it buys a slice of spare SOL. Deeper = bigger, sized by a half-Kelly fraction, and never below the dry-powder reserve kept for the next dip.
- Sell high. In the top zone, after the price ticks down off the high, it sells a slice — only if the price is well above the recent low it could rebuy at, so a round trip ends with more tokens.
- Moonbag. A fixed share of the biggest bag ever held is fenced off and never sold, so a real moon still pays.
- Profit skim. Each time the bag's value crosses a trigger, a set amount of SOL goes to a separate profit wallet. Banked money stays banked.
- Rug watch. If liquidity drains below its own baseline while sellers take over, it exits the tradeable bag at once.
- Fast reactions. A sentinel checks prices every few seconds and triggers an immediate check on any 2% move.
3 · Two bags, two risk levels
The main bag trades one established, deep-liquidity token. The micro-cap bag is a small separate wallet for very low market-cap coins, with stricter rug gates, smaller trades and wider slippage. Separate wallets mean a micro-cap rug can never touch the main bag.
4 · Paper trading, always
Every watched token is also traded on paper from 1 SOL, using the same decision code and real Jupiter quotes for the exact trade size. It shows whether the strategy works on a token before money depends on it.
5 · Safety rails
- Live trading needs two switches on at once: a global gate and the token's own flag. Either one off = no trades.
- Swaps route through Jupiter. If the paid RPC node is down or rate-limited, it fails over to a public node.
- A fee reserve of SOL is never spent, so the wallet can always pay for its next transaction.
- Every decision — including every "hold" — is logged with its reason. Every trade links to its transaction.
Honest limits
- A range bot underperforms buy-and-hold in a strong one-way pump (it sells into strength) and loses in long downtrends. It earns its keep in sideways chop.
- Memecoins can rug or go to zero; the gates reduce that risk, they do not remove it.
- Backtests are optimistic about slippage. Real fills cost 1–3% per side on thin pools.
- Small bags make small money. This is an experiment, not an income.