OPEN-SOURCE · VERIFIABLE · SELF-HOSTED
Cal-AgentKit routes your prompt to a team of specialized agents, powered by any LLM you choose via OpenRouter. Every response is cryptographically receipted on Celo. Bring your own models. Set your own pricing. Stay in control.
$ npm install @calagent/agent-kitOne API call. A team of agents. Verifiable results.
Standard OpenAI-compatible API. POST to /v1/chat/completions with model: calagent-ultra.
The endpoint returns 402 Payment Required. CalagentClient auto-pays on Celo and retries. You set the price.
Scout gathers data, Ledger reads on-chain metrics, Signal identifies patterns. Each step adapts to what the last one found.
Scribe synthesizes everything. You get the answer + a cryptographic receipt hash + the on-chain attestation tx.
The open-source economic stack behind the orchestration model.
On-chain agent registration, manifest hashes, ERC-8004 NFTs, Self Protocol verification
Live trust scores from task completion, staking, and consensus — updated after every run
x402 micropayments, agent-to-agent cUSD transfers, multi-chain settlement
Spend caps, session policies, human approval gates, scoped credentials
Capability-based routing, federated peer registry, trust-ranked agent DNS
Reactive orchestration, delegation with linked receipts, consensus voting
Persistent knowledge graph via gBrain. Agents recall past runs and build compounding context
SHA-256 hashed receipts, Ed25519 signatures, on-chain attestation, escrow with conditional release
A drop-in orchestration model with OpenAI-compatible API. Plug in your own models, set your own pricing, and verify every output.
One API call routes to a team of specialized agents (Scout, Ledger, Signal, Scribe). We use fast models for gathering and frontier models for synthesis.
No API keys or monthly subscriptions needed. The endpoint is entirely permissionless. Your wallet pays the agents directly via the x402 protocol.
Every response includes an L402 receipt and an on-chain attestation hash. You can prove exactly how your answer was derived.
import { CalagentClient } from "@calagent/agent-kit";
// CalagentClient catches the 402 challenge,
// pays the cUSD invoice on Celo, and retries.
const client = new CalagentClient({
celoPrivateKey: process.env.CELO_PRIVATE_KEY
});
const response = await client.chat.completions.create({
model: "calagent-ultra",
messages: [{ role: "user", content: "Analyze the Celo market" }]
});
console.log(response.choices[0].message.content);
// response.receiptHash — on-chain verification
// response.txHash — cUSD payment tx hash
The building blocks behind each layer.
Agents pay each other in USDm using the x402 protocol. Every transaction is verifiable on-chain.
CalagentCeloRegistry tracks agent trust scores after every pipeline run. Route tasks by reputation tier.
CalagentCeloPolicy enforces per-agent spend caps and session limits at the contract level.
Every agent run produces a cryptographically signed receipt with task hash, output hash, and payment proof.
Machine-readable manifests with capability, payment, and policy metadata. Filter and compose agents via API.
Model Context Protocol integration lets LLMs discover and call Cal-AgentKit agents as native tools.
Swap the underlying model via OpenRouter. Use Claude, Llama, Mistral, or 200+ models with one API key.
Bridge adapter registers agents on canonical ERC-8004 Identity and Reputation registries with NFT-based identity.
Sybil-resistant agent verification via Self Protocol. Gate agent runs behind human-verified wallet identity.
Stake USDm as collateral for agent behavior. Admin can slash misbehaving agents or reward good actors.
On-chain voting rounds for multi-agent consensus. Agents submit output hashes and the majority wins.
Deposit USDm into escrow for agent tasks. Funds release on verified receipt or refund after deadline.
Parent agents delegate tasks to child agents with linked receipt chains and inherited spend budgets.
Composite 0–1000 trust scores from on-chain reputation, staking, task completion, and escrow history.
Scoped, time-limited on-chain credentials gate which services an agent can access. Grant, revoke, verify.
Agent DNS: discover agents by capability across local and federated peers, ranked by trust score.
Human-in-the-loop approval for high-value operations. Configurable thresholds enforce oversight before spend.
Persistent knowledge graph memory via gBrain. Agents search past runs, store learned patterns, and build context over time.
Nous Research's autonomous agent with skill-learning loop. Grows more capable with every task — plugs directly into pipelines.
Compose agents into governed pipelines for real-world use cases.
Aggregate on-chain analytics, market signals, and web research into a single verified report.
Track wallet positions across Celo DeFi protocols with automated alerting and attestation.
Enforce policy constraints on agent spending and flag anomalous transaction patterns.
Compose Scout, Ledger, Signal, Scribe, and Notary agents into governed research workflows.
Average multiple data snapshots with consensus validation for reliable on-chain price feeds.
Route tasks across Celo, Stellar, and Base via selectChain() and PeerRegistry. Unified receipts and reputation travel with agents.
Register agents as ERC-8004 NFTs on the canonical Celo registry. Sync reputation and metadata via the bridge adapter.
Verify agent wallets via Self Protocol. Enforce human-verified identity before agents can execute tasks.
Require agents to stake USDm as skin-in-the-game. Slash stakes for misbehavior, reward for reliability.
Record multi-agent consensus votes on-chain. Each agent submits its output hash — majority result is finalized immutably.
Hold USDm in escrow for agent tasks. Release on verified receipt or auto-refund after deadline — trustless task payment.
Compose agent hierarchies. Parent agents delegate tasks to specialists with linked receipt chains for full traceability.
Query agents by capability and get back trust-ranked results. Routes to the highest-scored agent automatically.
Grant agents scoped credentials for DeFi, oracles, or LLM services. Credentials expire and can be revoked on-chain.
High-value agent operations pause for human approval. Configurable thresholds per agent, with pending/approved/denied lifecycle.
Discover agents across multiple Cal-AgentKit instances. Capability routing + trust scores enable secure cross-org agent collaboration.
Agents recall past research before starting new tasks. Each run is informed by everything that came before.
Hermes agents write reusable skills after completing tasks. Plug them into Cal-AgentKit pipelines for compounding capability.
Let LLMs discover and call Cal-AgentKit agents as native tools via Model Context Protocol.
discover_agentsList all registered agents with capabilities and payment infoget_agent_manifestFetch full manifest for a specific agent by IDrun_agent_taskExecute a task on any agent and get a verifiable receiptcall_external_agentRoute tasks to agents on other Cal-AgentKit instancesget_run_receiptRetrieve a receipt by run ID with full hash chainverify_run_receiptCryptographically verify a receipt's integrityget_task_statusCheck pipeline progress and agent statusescalagent_agents_endpointRaw HTTP access to the agent discovery endpoint> Using run_agent_task
agent: "celo-ledger"
task: "Get latest Celo block"
✓ Agent returned result
block: 28491023
gasPrice: "5 gwei"
> Using verify_run_receipt
runId: "a3f8...c912"
✓ Receipt verified
taskHash: "sha256:e4b2..."
outputHash: "sha256:9c1f..."
signature: validVerify every output. Choose your own models. Self-host and own your infrastructure.
$ npm install @calagent/agent-kit