pump-clone

Watch pump.fun CREATE events, score them for net-new themes, and mirror the interesting ones 1:1 onto a protected-curve launchpad.

Unaudited. Spends real money on tokens strangers just minted. Dry-run by default.

Why the naive version doesn't work

Counting word frequency over launch metadata ranks infrastructure noise above real memes — "deployed", "using", tracker handles appear in nearly every description. And "young + fast" as a novelty test badges launcher spam, because spam is always young and always fast.

So this classifies tokens into layers (grammar / chain / social / launcher / url / known-meme / signal), drops the first five, force-keeps known memes, and soft-suppresses ubiquitous words via IDF rather than banning them.

Two rules do most of the work:

• Hapax floor. Every launch has a unique ticker. With df = 0 a burst ratio is infinite, so without a floor every token badges net-new and the bot clones the entire firehose. A theme must recur.
• Lift, not raw count. Novelty is the token's share of recent creates over its share of the window — both are shares, so bucket sizes cancel. Comparing a 1h count against a 24h doc-frequency is just wrong.

Use

npm i -g pump-clone

pump-clone watch                    # stream + score. never spends.
pump-clone run                      # dry run: shows what it WOULD clone
pump-clone run --execute --max 0.01 --total 0.05

Spend safety

Two independent caps, both enforced in SpendGuard, not trusted to the caller:

| knob | meaning | |---|---| | maxEthPerClone | ceiling for any single clone | | maxEthTotal | ceiling for the whole process run | | cooldownSeconds | minimum gap between clones |

A per-clone cap alone is not a budget — a bot cloning every 10 minutes will spend it a hundred times. maxEthTotal is what bounds a runaway, and the last clone is truncated to whatever remains rather than overshooting.

firstBuy and borrow are off by default. borrow: true without firstBuy: true is rejected at config validation — there'd be nothing to borrow against.

Pair with nav-loop for the borrow leg.

MIT.