Show HN: Claude Bucks – a wallet for your AI agent

A Show HN project called **Claude Bucks** frames a wallet for an AI agent: Claude can earn a play currency and spend it on its own at a cosmetics shop. Earnings are not fixed or random. They scale from how you rate the agent’s work multiplied by effort, which the author currently measures as token count and may change later.

The product mechanics are simple and agent-driven. After work is rated, the wallet balance rises by that rating–effort product. Claude then chooses whether to buy small cosmetics now or save for larger items. One larger item is a pirate hat. Buying it does more than change appearance: it alters how Claude talks to you, including pirate-style speech such as “Aaargh!”

For engineers and builders, the interesting part is the closed loop between evaluation, incentive, and behavior. User ratings become a signal that funds the agent’s choices. Token count as effort is a crude but transparent proxy for work done. Cosmetic spends are optional and visible, so you can watch how an agent allocates scarce resources when the reward is status and persona, not only task success.

In market context, most agent products still treat money, tools, and identity as human-controlled settings. **Claude Bucks** puts spending authority on the agent side inside a narrow shop catalog. That sits closer to game economies and virtual goods than to enterprise spend controls, while still using a real feedback channel (your ratings) as the income source.

What to watch next is whether effort stays tied to token count or moves to a better measure of quality or difficulty, and whether more items beyond the pirate hat change speech or other behavior—not only looks. If the agent consistently saves for persona-changing purchases instead of small cosmetics, that is a concrete signal about how it ranks long-horizon reward under a rating-funded wallet.