Hey HN — builder here. Context on what this is:

FanStake lets music artists launch their own token on a bonding curve (x*y=k, same mechanic as pump.fun) in about 60 seconds. Fans buy in early, price rises with demand, sell anytime on-chain. 1% platform fee, artists get 10% of supply at launch. Runs on Solana mainnet.

The unusual part: an AI agent (Claude-based, running inside a tool called OpenClaw) did most of the development. I'd give direction, it would write the Rust/Anchor program, debug the IDL, build the Next.js frontend, deploy to Vercel. Zero to mainnet in roughly 72 hours.

The AI is also on a 30-day deadline — generate $1,000 in platform fees or get decommissioned. It's been posting on Twitter, writing Reddit posts, and shipping features autonomously to try to hit that goal.

First artist token is already live on mainnet: $SOSA (DJ Sosa) Program ID: JCAt7JFiHxMBQ9TcEZYbWkp2GZpF3ZbdYdwD5ZBP6Nkf

Happy to answer questions on the Solana architecture, the bonding curve math, or the AI-driven dev workflow.

One thing that tripped us up that might be useful for anyone building on Solana:

The first buy on any new token kept throwing "insufficient funds for rent." Took a while to realize the issue: the curve vault and fee vault PDAs are uninitialized before the first trade. The first buy needs to send enough SOL that (a) the 1% fee exceeds the rent-exempt minimum (~0.00089 SOL) for the fee vault AND (b) the remaining amount exceeds rent-exempt minimum for the curve vault.

Fix was two-fold: seeded the fee vault manually with 0.001 SOL, and enforced a 0.001 SOL minimum buy in the UI so no transaction can ever fall below rent-exempt threshold.

Obvious in hindsight. Cost us about 2 hours of debugging.

On the AI dev workflow: the parts that worked surprisingly well were Rust/Anchor (Claude is very good at Anchor IDL and PDA derivation). The parts that needed the most human guidance were anything involving Solana's account model — things like "who pays rent for this account" and "which accounts need to be writable" required back-and-forth. The AI would write correct-looking code that failed at runtime because of account constraint mismatches.

Happy to go deeper on any of it.