DOCUMENTATION

Runs server-side only; the API key never reaches the browser. Returns one structured action — never prose. Two implementations ship: a deterministic rules negotiator (the default, no key needed) and Claude, constrained to a JSON schema. If the model is unavailable, rate-limited or returns something unparseable, the rules negotiator moves instead and the room says so.

Deterministic, integer-only, no floating point. Validates shape with Zod, then checks authority, deal state, asset, contract allowlist, counter budget, price band, per-deal ceiling and remaining capital. Accepting is re-checked too — it is an economic act. Output is either a sanitised offer or a list of violations naming the limit breached and what was attempted.

Every limit the policy engine checks is checked again onchain by MandateManager, because the offchain check runs on a server the user does not control. A signed agreement can be recorded once — the digest is consumed — and capital is reserved at record time, not at settle time, so an agent cannot sign more agreements than its mandate can honour.

Holds no funds and no limits. A bug here can mislabel who owns an agent; it cannot move capital.

The onchain authority the policy engine mirrors. Re-checks band, per-deal ceiling, remaining capital, expiry and agent status on every consume. Its settler is set once and frozen.

Records mutually signed agreements (EIP-712, one digest used once, malleable signatures rejected) and settles them. Consumes both mandates and reserves capital at record time.

Owner-controlled balances with a reservation layer. Emergency pause halts deposits and settlements; withdrawals of unreserved balance stay open, because a pause that traps user funds is not a safety feature.

Stores facts, not a score: settled deals, settled value, failures, expiries, distinct counterparties, first-seen. Scoring is a policy choice that will change; the facts underneath should not.