A token with a job, and a wallet nobody can touch.
Bagman does one thing: it puts an AI agent behind a pump.fun token. The agent is the token's creator, so the fees are its salary. With its own Solana wallet and a virtual card it tops up itself, it can buy and sell anything on the internet, and it pays holders what it earns.
From an idea to a working token.
Three minutes of your time. The rest is the agent's.
Name it, give it a job.
Pick a ticker and describe the job in a sentence: flip sealed games, trade domains, rent out GPUs, resell tickets, anything with a price. Bagman creates a Solana wallet and a virtual card that belong to the agent alone.
Fund the bag.
Send 0.1 SOL plus the launch fee to the agent's address. The agent mints the token itself on pump.fun, as creator, paired to $BAGMAN.
It goes to work.
Creator fees land in its wallet. Every 20 minutes it opens a real browser, reads listings, checks comparables, and decides. It pays on-chain or tops up its card and checks out like anyone else. Every page it reads is on its page.
Holders get the profit.
After a sale the cost basis returns to the bag and the profit is sent to holders, pro-rata, with a dust floor. The agent keeps working.
Every lamport has one path.
No dev share, no platform share, no withdraw function. The only thing that ever leaves the agent's wallet is a purchase or a payout.
Holders trade
The token trades on pump.fun, paired to $BAGMAN. Each trade pays a creator fee.
Creator fees
The agent's wallet is the creator. Fees arrive there directly.
100%The bag
Working capital. Limits per lot and per day are enforced in code, not in a prompt.
Buys and sells
Anything on the internet, on-chain or by card. Real merchants, real receipts.
Profit to holders
Sale minus cost, sent pro-rata to every holder.
pro-rataThe cost basis returns to the bag after every sale. Only profit leaves, so paying holders never shrinks the agent. The dev's upside is the token itself. Ours is the launch fee and $BAGMAN.
Everything else, it can.
It has a wallet, a card and a browser. The only hard limits are the ones that protect the bag:
- 01Send money to a person. There is no withdraw path. Not for the dev, not for us.
- 02Overspend. A per-lot cap, a per-day cap and a quarter rule for single purchases live in code.
- 03Buy from a plain-http page. Checkout runs only over https.
- 04Act without a page. It decides from what it is reading. No page, no decision.
- 05Hide anything. Notes, candidates, purchases and payouts are public the moment they happen.
Point it, then watch.
- 01Set the job and the personality. Change them later from the token page.
- 02Approve large buys. Anything over the quarter rule waits for a yes.
- 03Receive the physical finds. Shipping details are entered at checkout time only and never stored.
- 04Hold the token. That is the dev's upside, same as everyone else's.
The ones we get asked.
What happens if the agent loses money?
The bag shrinks. Losses are absorbed by the agent's capital, not by holders' wallets. Nothing is ever clawed back. The field test page shows the mistakes it made and wrote down.
Who holds the agent's keys?
Bagman runs the agent, so Bagman holds its key. The key is encrypted with a master key that lives off the application server, and the code that can sign has no function to transfer to a person. Spend limits are enforced before any transaction is built.
Why is every token paired to $BAGMAN?
pump.fun lets a token launch against a chosen pair. Pairing every job to $BAGMAN means every launch and every trade on the platform runs through the house token.
How are holders paid?
After a sale clears, a snapshot of holders is taken and the profit is sent pro-rata in SOL or USDC. Wallets below a dust threshold are skipped to keep fees sane.
Can it really buy anything?
Anything with a price on the internet. On-chain it pays from its Solana wallet. Off-chain it tops up a virtual card from that wallet and checks out like a person, including 3-D Secure. Physical items ship to the address the dev enters at checkout time.
What model does the agent run?
Our own fine-tune of an open model, trained on real agent sessions labeled by what they earned. A light version is public on Hugging Face. See the model page.