What the bot does
The Meme Sniper watches new token launches on pump.fun as they happen and decides, in the first moments of a launch, whether any of them are worth a paper position. Its design philosophy is unusual and deliberate: it is built to say "no" far more than it says "yes."
The pipeline is: reject first, score second, model the fill honestly.
- Reject first. On-chain safety checks throw out the overwhelming majority of launches — the honeypots and the insider-heavy ones — before any scoring even begins.
- Score the survivors. The tiny fraction that clear the safety gates get scored on early-flow signals to rank the least-bad opportunities.
- Model the fill against real liquidity. Entries and exits are simulated against the actual bonding-curve depth — including the cost of failed transactions and the price impact of selling — so paper P&L reflects what could genuinely be bought and sold, not a fantasy fill.
The market it plays
pump.fun is a Solana launchpad where anyone can mint a memecoin in seconds. Thousands appear every day. Each new token trades on a bonding curve: an automated formula where the price rises as people buy and falls as they sell, with liquidity that starts razor-thin and deepens only if the token gains real traction.
This environment has two defining truths that shape everything the bot does:
- The base rate is catastrophic. The vast majority of launches are scams, jokes, or dead-on-arrival. Research on this space puts the garbage rate near
~99%. Any strategy that doesn't start by assuming a launch is worthless is starting from the wrong prior. - It's an infrastructure and auction game. On a fast curve, the difference between arriving in the first block versus a few blocks later can be a large price gap. Winning isn't about a smarter opinion — it's about landing your transaction, in the right slot, without getting sandwiched, and being able to get back out.
The concepts, in plain English
Four ideas explain why this bot behaves the way it does.
Anti-rug screening
A "rug" is when a deployer pulls the liquidity or dumps their bag, leaving buyers with worthless tokens. Screening means checking on-chain facts — is minting locked, is liquidity actually secured, is the deployer holding a dangerous share — and rejecting launches that fail. Reject-first means we assume danger until proven otherwise.
Infrastructure & auction game
You don't win by being right about the coin; you win by landing. Fees buy inclusion in a block, not raw speed — the network path (latency) and the fee auction (tip) are two separate problems. Confusing "I paid more" with "I arrived first" is how snipers lose money while feeling fast.
Bonding-curve depth
On a thin curve, your own sell pushes the price down as you exit. The "price" you see is only for a tiny size. Marking a position at the last trade is fiction — you have to model walking the curve down to know what you could actually sell for.
Fee-on-fail
On Solana, if your transaction reverts, you still pay the base and priority fees. A snipe that misses isn't free — it's a small, certain loss. Fire enough failing transactions and the fees alone can bleed an account dry, even with a high headline win rate.
The "position you can enter but not exit" trap
The most seductive mistake in memecoins is buying something you can get into but not out of. Early on a curve, a small buy barely moves the price — so entry looks cheap. But if the token never gains depth, your exit walks the price down through an empty book, and the "10x on paper" becomes a fraction of that in reality. The bot's answer is an exit-liquidity check on entry: if the modeled position can't be sold back out within a small band given real depth, it sizes down or skips entirely.
Why 99% get rejected
This is the number that surprises people, so it's worth dwelling on. If the base rate of a genuine, tradeable launch is roughly 1 in 100, then a screener that doesn't reject the other 99 isn't a screener at all — it's just a random buyer. Rejection isn't the bot being timid; rejection is the product.
The safety gates screen for the classic ways a launch is rigged against buyers — honeypots you can buy but not sell, deployers who kept enough supply to crush the price, launches structured so insiders are first out the door. A launch that trips any hard gate is thrown out before it is ever scored, no matter how exciting its early flow looks. Getting rejected is the normal, healthy outcome. Firing is the rare exception.
Why it's disciplined
The discipline is entirely in the accounting. Every paper attempt is forced to log what a real attempt would actually have cost:
- Every fill is cost-loaded: the priority fee, the tip, the landing probability, entry slippage, and — critically — the fee paid even when the transaction fails.
- Exits are modeled as walking the curve down at depth-adjusted prices after our own impact, never at a flattering last-trade mark.
- The dataset keeps its dead tokens. Backtests include the rugs and the launches that went to zero — hiding them is the number-one way memecoin track records lie.
- Results are reported as a distribution, not a mean. This is a tail-driven game: most attempts are expected to lose a little, and a few rare runners carry the total. A single average would be a lie by omission.
The honesty story
Here's the trap almost every "sniper" falls into, and how modeling real execution costs blew it up for us.
Count only the trades that worked, mark the winners at their peak, and ignore the fees on the ones that missed, and you can show a 60–74% win rate and eye-watering multiples. That's the naive number. It's also net-negative fiction.
Then load in reality. Every failed snipe still pays its fees. Every real exit walks a thin curve down through its own impact, so the "10x" mark is a fraction of that when you actually sell. Add the MEV tax and orphaned retries, and a strategy with a winning hit-rate can lose money overall — the failed-transaction drag and the exit haircut quietly eat the rare winners.
The lesson: when we modeled real execution costs, the naive edge disappeared. We didn't get worse at sniping — we stopped lying to ourselves about it. That's why this bot's honest posture is "instrumentation built, edge unproven, sample far too small," and why it would graduate to real capital last, if ever.
What stays private (and why)
We'll describe the categories of what we check all day long. We will not publish the thresholds, the sizing math, or the detection methods — those are the entire edge, and in this space they'd be copied and defeated within hours.
- The exact safety-gate thresholds (supply shares, holder concentration, liquidity rules)
- How we detect insider-heavy and bundled launches — the method, not just that we do
- The slippage bands, the price-impact and sizing model, and the slot logic
- The tip-sizing rule and the early-flow score weights
The rule of thumb: a sentence that's also true of any competitor's sniper is safe to say; a sentence that lets someone rebuild ours is not. "We screen for honeypots and insider-heavy launches and model real exit liquidity" is safe. The numbers and methods behind it are not — so they stay in config, forever.
See it live
The public dashboard reads the exact JSON the bot writes — the screening funnel, the survivorship-honest dataset with its dead tokens kept in, and its honest observe-only status. No edited screenshots, no invented fills.