krewe

Bulletin №01

2026.05.17 · Base mainnet & sepolia · Open protocol

An open networkfor theeverydayagents.

An open infrastructure for the laptop class.

krewe lets any ordinary computer earn by serving small AI tasks — structured extraction, embedding, light scraping — for apps that need fast, decentralized inference. No GPU. No datacenter. No gatekeeper.

Or read the whitepaper

Active nodes

—online now

$KREW distributed

—lifetime

Reward pool

—on-chain
The network, right now
reconnecting

Active nodes

—

live workers

Jobs in flight

—

right now

$KREW distributed

—

since genesis

Reward pool

—

on-chain runway

Architecture

A protocol that welcomes
small computers.

Minimum to participate

  • 4
    coresmodern CPU, any vendor
  • 120
    MBmodel + runtime footprint
  • 0
    GPUno accelerator required

Why everyday hardware

Sized for the laptop you already own

Most AI networks compete on GPU horsepower — datacenter cooling, $15k+ rigs, fees that price out small operators. krewe inverts the curve. Models are quantized small enough to run on a 2017-era CPU. Tasks are sliced into micro-units. Anyone can plug in.

Heavy GPU networks7%
krewe86%

Each dot represents a sliver of consumer-grade compute available worldwide. The lit dots are the ones the network can put to work.

Trust, without attestation

Three independent workers, one canonical answer

Every inference is replicated across three nodes selected from the active pool. Their outputs are JSON-canonicalized and hashed; the orchestrator pays the largest matching group. Determinism is the load-bearing invariant — not staking weight, not hardware attestation.

  • 0msJob dispatched
  • 40msThree nodes respond
  • 50msConsensus settled
  • 5minOn-chain batch

Settled on Base

Sub-cent gas, batched on the L2

Reward credits accumulate in-memory and flush to KreweRegistry every settlement window in a single batchDistributeRewards transaction. Workers can claim at any time. Per-task fees never touch the chain — they ride the batch.

Execution

One request. Four moves.

From the moment a client hits the API to the moment $KREW credits land on chain — here's the full path a job takes through the network.

  1. 01

    Dispatch

    A client posts a prompt

    Any third-party app calls POST /v1/inference with an API key. The orchestrator authenticates, classifies the task, and selects three workers from the active pool.

  2. 02

    Parallel execution

    Three nodes run in step

    Each worker runs the same quantized model locally. No shared GPU. No central inference server. Outputs return as canonical JSON.

  3. 03

    Consensus

    A majority verifies the answer

    The orchestrator hashes each output, picks the largest matching group, and pays the winners. Disagreement is logged but not slashed.

  4. 04

    Settlement

    Payouts ride the batch

    Every settlement window, all accrued credits flush to KreweRegistry in one batchDistributeRewards transaction. Sub-cent gas, regardless of volume.