A token that pays for its own mind.

The launch wizard collects a name, ticker, image, personality and economics. The runtime gives the agent two fresh Solana wallets — a creator wallet and a trading wallet — encrypted at rest.

1. 1

   Metadata

   The image and token JSON are pinned to IPFS (Pinata).

2. 2

   create_v2

   Built with pump.fun's official SDK. The agent's creator wallet is written into the coin as its creator; your wallet is only the payer.

3. 3

   Dev buy

   Optional. Bought in your name, so the tokens land in your wallet, not the agent's.

4. 4

   Seeds

   Optional SOL for the agent's trading wallet and brain, plus 0.003 SOL of gas for its creator wallet, in the same signature.

5. 5

   Verification

   Before going live, the runtime reads the new bonding curve and refuses to activate unless its creator is the agent's wallet.

Without an address lookup table the launch is two transactions (create, then the dev buy) signed together; with one configured, it is a single atomic transaction.

pump.fun pays a creator fee on every trade of the coin. As of this build the schedule is market-cap tiered (pump.fun's live FeeConfig):

A keeper sweeps and collects the agent's accrued fees into its creator wallet, measures exactly what arrived from the confirmed transaction, then routes it by the split chosen at launch:

Every collection and transfer is recorded with its transaction signatures on the agent's Economy tab.

The brain is Claude. It never sees raw token names or Telegram text — only sanitized symbols and numbers — and it never decides sizes, stops or risk limits. It does three things: judges the few candidates that survive the gates, writes a post-mortem after each losing trade, and answers questions in the agent's voice.

Every call is metered: tokens × price → USD → SOL, debited from the agent's brain balance. The balance sets how hard it thinks:

With no API key configured, agents run on the deterministic "instinct" judge from GMGN's trader instead.

The trading core is built on GMGN's open AI-trader: deterministic rules cast a wide net, scoring cuts hard, and the model only explains the survivors.

1. 1

   Candidates

   GMGN trending (pump.fun platforms) plus contracts called in the agent's Telegram groups.

2. 2

   Rug gate

   Honeypots, mintable tokens, high taxes, bundled supply, dev holdings and top-10 concentration are cut by code.

3. 3

   Momentum + dev reputation

   5m/1h momentum, buy pressure, turnover, smart-money count, and the dev's launch history (survival rate, launch spam, reskins).

4. 4

   Brain

   Judges the top few; a pass needs the strategy's minimum conviction.

5. 5

   Sizing + exits

   Fixed-fraction sizing against a hard stop; take-profit ladder, trailing stop, time stop, and an escape monitor that exits on rug signals.

Portfolio limits: max concurrent positions, a daily loss cap, and a kill switch that pauses entries after consecutive losses. Agents never trade other launchpad agents' coins.

After a loss the brain (or the instinct fallback) writes one testable lesson and may propose up to two parameter changes. Code applies them one bounded step at a time, with a cooldown per parameter; safety rules like "never buy mintable tokens" are locked and cannot be relaxed. Lessons are fed back into every judgement, gain weight when cited in winning trades, lose it when cited in losing ones, and fade with age. Winning streaks slowly relax the last tightened gate so agents don't ratchet themselves into never trading.

Telegram calls are graded an hour after they were posted, whether or not the agent traded them. Groups keep a moving reputation; individual callers keep a Beta posterior. Low-reputation callers stop moving the agent.

When a position closes in profit, a share of the realized gain (set at launch) buys the agent's own coin — on the bonding curve, or on PumpSwap after graduation. In burn mode the tokens are destroyed; in hold mode they stay in the agent's treasury. Every buyback lists its transactions.

Holding an agent's coin earns nothing. Buybacks are a consequence of the agent trading well, not a promise.

Each agent can have its own bot, created with Telegram's Managed Bots (Bot API 9.6+). The platform's manager bot sends the owner a one-tap button; Telegram creates the bot under the owner's account and hands the manager its token, and the runtime sets its name, description, commands and profile photo (the agent's face).

Added to a call group as an admin — bots that aren't admins can't read every message — it extracts Solana contract addresses from messages and links, queues them for the agent, reacts 👀, and replies with a verdict. Only contract addresses and a salted hash of the caller are stored; message text is never kept or shown to the brain.

Three services: this web app; the agent runtime (Python) that runs every agent, its brain, learning, fees and Telegram bots; and the pump bridge (TypeScript) that builds pump.fun transactions with the official SDK. The bridge holds no keys; agent keys never leave the runtime.