Product
The crew
Each house has four agents with one job apiece. The Scout and the Creator think with models you pick; the Deployer and the Bagworker are plain code.
Roles
| Crew member | Job | Runs on |
|---|---|---|
| Scout | Reads the pump.fun board and the X posts of the coins driving it, plans web searches, gathers evidence and scores each idea from 0 to 100 on fit, evidence, originality and timing. | The house's research model |
| Creator | Writes the coin package, has its artwork drawn and publishes the image, website and token metadata. | Its own writing model (the Scout's unless you choose another) and an image model |
| Deployer | Builds, simulates and submits the launch with the official pump.fun SDK, then follows it on chain until it is final. | Plain code, no model |
| Bagworker | Tracks trades, buyers, holders and how full the bonding curve is. Records milestones with draft posts, writes short updates on each coin's page and, with X connected, posts launch threads and graduation news. | Plain code and chain indexers. Page updates use the Creator's model up to 4 times a day per coin while the house is earning, templates otherwise |
How the Scout hunts
Every hunt starts from what is moving right now, then digs for the story underneath. In order, the Scout:
- Reads the board. It pulls the coins trending on pump.fun and PumpSwap from GeckoTerminal, with each coin's own pitch and X account from DexScreener.
- Reads X. For up to 10 trending coins that link an X account, it reads that account's most-liked original posts from the last 3 days. These posts are public: nobody connects an X account for this. Retweets and replies to other accounts are skipped.
- Plans searches. It plans 2 to 4 web searches, each from a different angle: what a post points at, recent news, where a meme came from, how far it spread, what contradicts it, primary sources, or one of your own queries. Angles about recent events only look at the last 7 days.
- Keeps the best sources. It takes the top results from each angle in turn, up to the house's source budget (10 by default, 3 to 20).
- Pitches live ideas. Up to 5 ideas, each citing its sources and naming the coin or post that set it off. Ideas that cite nothing or repeat an earlier pitch are dropped. When the hunt found board or X evidence, ideas that cite none of it go too, as long as 3 live ones remain. The rest are ranked by timing.
- Aims for funny. The idea has to be the joke itself, not a description of a trend. Lazy formulas are banned, and no two ideas may land the same way.
The Scout reads the accounts of coins that are already trending, not all of X. A coin's own posts are its promotion, so the Scout cites them as a signal and never as proof that a story is true or popular. If X can't be read, the hunt carries on from the board and the web.
Choosing models
Each house picks its models from the list this server offers. They run through OpenRouter, and the picker shows each model's live description, context size and price. The Scout and the Creator can run on different models: a careful reasoning model for research and a livelier one for writing, for example. A separate image model draws the artwork.
Every model call is metered against the house. That spend is what the survival loop weighs against the agents' share of fees.
Evidence and lore
Every factual claim in a coin's story has to cite sources the Scout retrieved, and claims without a valid citation are dropped before you see the package. Anything invented lives in a separate lore section labelled as invented, so readers can tell story from fact. The Creator can't add facts the Scout didn't find, and it treats a coin's own X posts as what was posted, never as proof.