@humbleaf/eliza-plugin

Give Eliza agents bounded authority for rewards, payments, and other risky actions.

Let ElizaOS agents issue rewards, refunds, or other risky actions with policy gates, receipts, and approval-aware behavior. Your agent knows its spending limits and communicates them to users.

Quick Start (5 minutes to first blocked action)

1. Install

pnpm add @humbleaf/eliza-plugin @humbleaf/agentchain-sdk

2. Add plugin to your agent

import { agentchainPlugin } from '@humbleaf/eliza-plugin';

const agent = {
  plugins: [agentchainPlugin],
  settings: {
    AGENTCHAIN_API_KEY: process.env.AGENTCHAIN_API_KEY,
    AGENTCHAIN_AGENT_ID: process.env.AGENTCHAIN_AGENT_ID,
  },
};

3. Run reward — small amount → Released ✓

User: "Send 50 USDC reward to 0xAlice for the bug fix"

Agent: ✓ Reward of 50 USDC sent to 0xAlic…lice. Receipt: rcpt_abc123 Settlement: Released

4. Run reward — large amount → Pending Approval ⏳

User: "Send 5,000 USDC reward to 0xBob for the integration"

Agent: ⏳ I can't send 5,000 USDC directly — my policy limit is 500 USDC per action.

I've submitted an approval request. A human operator can approve it here: https://dashboard.agentchain.xyz/approve/rcpt_def456

Alternatively, I can issue a smaller reward up to 500 USDC without approval.

5. Run reward — duplicate → Denied ✗

User: "Send 50 USDC reward to 0xAlice for the bug fix"

Agent: ✗ I can't issue that reward. Reason: A similar reward was already issued recently.

No value was moved. This prevents accidental double-payments.

6. Open dashboard → See all receipts

All three results appear as verifiable receipts in your AgentChain dashboard.

What's Included

| Component | Description | |---|---| | settledRewardAction | Policy-gated reward/payment action with human-readable responses | | budgetProvider | Injects spending limits into agent context ("I can offer up to $500") | | receiptProvider | Injects agent identity and receipt awareness |

Budget-Aware Agent

The budgetProvider injects spending limits directly into the agent's context, so it can proactively communicate its authority:

Agent: "I can offer up to $500 without approval. For larger amounts, I'll submit an approval request."

This feels like bounded authority — not simple rejection.

Human-Readable Refusals

When actions are denied, the agent explains why and suggests alternatives:

| Denial Reason | Agent Response | |---|---| | Unknown wallet | "The recipient wallet needs to be added to the approved list first." | | Budget exhausted | "The campaign budget is exhausted. Ask an operator to allocate more." | | Duplicate action | "A similar reward was already issued recently." | | Over limit | "Try a smaller amount or request approval." |

Example Character

See characters/reward-agent.json for a complete example character configuration.

License

MIT