An open network for everyday agents.

The thesis is simple. The majority of the world's compute capacity isn't in datacenters — it's sitting idle on personal devices. When a task can be split small enough, those devices become a credible alternative to centralized inference vendors.

krewe makes that possible through three load-bearing design choices:

Workers run small, deterministic models locally — embedding pipelines, structured extraction, light scraping. The model footprint is ~120 MB; execution requires nothing more than a 2017-era CPU. No GPU, no accelerator, no datacenter colocation.

Every inference is replicated across three workers selected from the active pool. Their outputs are canonicalized as JSON and hashed; the orchestrator pays the largest matching group. Trust is enforced by consensus over reproducible output — not by staking weight, not by hardware attestation.

Disagreement is logged and degrades the offending worker's score, but it's never slashed. The penalty is implicit: lower-scored workers receive fewer jobs.

Reward credits accrue in-memory and flush to the on-chain registry every settlement window in a single batchDistributeRewards transaction. Per-task fees never touch the chain — they ride the batch. Sub-cent gas costs, regardless of inference volume.

$KREW is a fixed-supply ERC-20 (1B cap, 18 decimals, burnable, EIP-2612 permit). The full supply was minted to the treasury at genesis; no further mints are possible. The contract has no privileged functions beyond ownership renunciation.

Reward pool inflows are governed by the protocol multisig. Worker stake is segregated from the reward pool — emergency admin withdrawals cannot touch collateral.

The MVP is live on Base Sepolia. The path to mainnet is admin/orator rotation to a multisig, per-client API authentication, low-watermark reward-pool alarms, and adversarial determinism testing for the embedding and agent task kinds.