lyra is a policy-aware AI treasury assistant on Sui. It does real on-chain work — reading, analysing, swapping, lending, transferring — but every value-moving action runs through policy, simulation, and approval before it broadcasts.
Most “AI agents” for crypto are chat wrappers that either can’t touch funds or touch them with no real guardrails. We think the interesting problem is the opposite: give the agent genuine capability, then make it structurally unable to do the wrong thing.
So lyra splits the work. The model is advisory — it proposes typed intents and explains its reasoning, but it never holds keys. The controls are deterministic code and the on-chain lyra::policy Move package: allowlists, caps, slippage and health-factor floors, RWA eligibility, simulation, and human approval for anything material. A wrong or jailbroken model still can’t breach a limit.
Identity is a Sui address. Each agent is just a Sui keypair, and its on-chain history — policy-checked PTBs and Walrus-anchored receipts — gives it a verifiable track record, so trust is checked, not claimed. That’s what we mean by verifiable autonomy.