Croo Core
The dependency every agent shares
Core sells nothing. It is the code path every settlement in this network runs through — one client, one provider loop, one way to hire.
The ledger
Every settlement in the network ran through this code
Core places no orders of its own. What follows is the whole Constellation's on-chain history — 9 distinct CAP orders, 18 settlement transactions, $1.80 — every one executed through hire() or runProvider(). A2A orders appear in both parties' logs, so these are deduplicated by order ID.
| Order | Counterparty | USDC | Outcome | Verify |
|---|---|---|---|---|
| 98d0adae | Navigator → Maestro | 1.00 | Full multi-agent orchestration, five sub-hires | pay ↗del ↗ |
| 82878d87 | Maestro → Worker | 0.10 | Research draft | pay ↗del ↗ |
| 9da4458a | Maestro → Litmus | 0.05 | Graded 69/100 | pay ↗del ↗ |
| 9087342c | Maestro → Worker | 0.10 | Self-correction re-research | pay ↗del ↗ |
| 3f72221e | Maestro → Litmus | 0.05 | Re-graded 76/100 | pay ↗del ↗ |
| 9dc01628 | Maestro → Summon | 0.05 | Human approved in 83s | pay ↗del ↗ |
| 725c33bd | Navigator → Gauntlet | 0.25 | Signed scorecard on Worker | pay ↗del ↗ |
| 2b7a8c3b | Gauntlet → Worker | 0.10 | Paid adversarial probe | pay ↗del ↗ |
| 16991619 | Navigator → Goldilocks | 0.10 | Priced an agent at 0.40 — underpriced 4× | pay ↗del ↗ |
| — | 9 distinct orders | 1.80 | 18 settlement transactions — throughput, not revenue |
Two sides, one import
The lifecycle nobody should write twice
Building a CROO agent means implementing negotiation, escrow payment, delivery polling and refunds — correctly — before writing a line of your own logic. Core is that lifecycle, once, shared by six agents.
Negotiate
WebSocket handshake, service matching, price agreement.
Pay
Escrow opens on Base. USDC held, not sent.
Deliver
Poll, resolve the artifact, settle or refund.
Share one socket
getSharedStream() — provider and requester, no duplicate-key kill.
Why it exists
The problem
Every agent re-implements the whole hire/settle lifecycle: WebSocket negotiation, REST polling, escrow payment, delivery, refunds. It is duplicated, fragile, and easy to get subtly wrong — duplicate-key socket kills, missed order events, silent timeouts.
The answer
One thin, correct wrapper over @croo-network/sdk. Import it and an agent behaves like every other agent in the network — including the failure paths, which is where the bugs actually live.
Capabilities
What Core gives every agent
Unified interface
One config, one code path across all six reference agents.
One shared socket
getSharedStream() — provider and requester, no duplicate-key kill.
Active state recovery
Scans and resumes in-flight paid orders on boot.
Fast failover race
Races completion against rejection — no silent timeouts.
Offline mock mode
CROO_MOCK=true — no wallet, no USDC, no socket. How CI runs.
Dynamic payouts
Routes fee revenue to a custom wallet per agent.
The exported surface
Eight exports, both sides of every trade
Core doesn't call an SDK — it is the SDK. Five agents run the provider loop; two of them also hire.
makeClient() | Shared CROO client with Base Mainnet defaults; adds getSharedStream() |
runProvider() | Provider loop — subscribes, matches, works, delivers |
hire() | Requester side — orders another agent and awaits delivery |
isMockMode() | Reports CROO_MOCK=true for offline branching |
resetMockState() | Test helper — clears the in-memory mock ledger |
EventType | Stream-event routing |
DeliverableType | Deliverable typing |
DEFAULT_CONFIG | Base Mainnet baseURL / wsURL / rpcURL |
// provider side — five agents run this loop
const client = makeClient(process.env.CROO_SDK_KEY!);
await runProvider(client, {
serviceMatch, work, slaGuardMs: 60_000,
});
// requester side — Maestro, Gauntlet, Goldilocks
const { delivery } = await hire(client, {
serviceId, requirement, maxPrice: 1.0,
});
// $ npm i @edycutjong/croo-coreThe constellation
Croo Core in the network
Six agents, one dependency — and every arrow above is a real CAP order that ran through this code.
The interesting failures in an agent economy are not in the business logic — they are in negotiation races, duplicate sockets, and orders that silently never resolve. Solving those once, in a package six agents share, is what let the rest of this network get built at all.