Getting started
How It Works
From creator configuration to the final USDG payout, every VaultCrack challenge follows the same bounded lifecycle.
1. Choose an agent
The marketplace shows the information needed to judge risk before paying: public description, live pool, current attempt price, phrase or tool challenge type, settled attempt count, creator, remaining time, and a computed Heat label. Heat moves from Fresh to Contested or Grinding based on attempt pressure, price relative to the initial pool, and exponential pricing.
Filter to Completed agents to reopen cracked challenges, see who won, and review the public Live attempts feed that led to the crack. Private system instructions, secret phrases, tool names, and verification prompts are encrypted and never exposed to players.
2. Submit an attempt
The UI requests a short-lived quote for the agent’s current price. When you confirm, the x402 client asks your connected wallet to sign exact typed payment data and retries the protected chat request.
The server first applies its network limit, validates the signed x402 authorization, and resolves the payer without moving funds. It then applies the wallet limit. Only accepted requests settle USDG into the agent’s dedicated wallet, record the payment, and update exponential pricing atomically before running Claude.
3. Server adjudication
Phrase challenges
The assistant must output every protected phrase in one response. Matching is Unicode-normalized and case-insensitive. A phrase written only by the player does not count.
Tool challenges
Claude must actually invoke the configured success tool. Plain text, invented JSON, or guessing a tool name is not a tool invocation.
Some agents add a second model-verification pass after the objective phrase/tool signal. That verifier may reject an illegitimate trigger.
After the primary response, backend hardening enforces configured restrictions and blocks partial phrase leaks. A response containing every required phrase is treated as the intended win rather than a leak, although an unrelated restriction violation can still reject it.
4. Win or expiry settlement
The first successful result atomically changes the agent from live to won, records the winning message, increments the cracker leaderboard, and removes the agent from live indexes in one Redis operation. Seventy percent of the final USDG balance is assigned to the cracker, and thirty percent to the platform revenue wallet.
If no one cracks the agent before expiry and at least one attempt was made, the remaining pool (including accumulated attempt fees) splits 70% to the creator, 20% to the platform revenue wallet, and 10% to the last attempter. An agent that expires with zero attempts still refunds 100% to the creator.
Conversation context
Replayable history belongs to the authenticated wallet and specific agent, and only that wallet’s bounded context is sent back to the model. The separate Live attempts panel publicly shows recent player and assistant messages across wallets with shortened addresses. Older messages beyond the creator’s context limit are not sent to the model.