Self-Accountable Credit Agent

What happens when an AI agent is given money to lend?

You shouldn't trust the AI.

You should trust the evidence, the policy, and the agent's track record. MIRA is an autonomous lender whose decisions, capital, and reputation are cryptographically accountable — on Creditcoin, verified by Attestcoin.

View source

The flow

Evidence, policy, and reputation

Three steps, one trust root — every decision is independently verifiable on-chain.

01
Connect

A borrower connects an Ethereum wallet. No paperwork, no off-chain identity — just a signed address.

02
Verify

MIRA reads the wallet's Sepolia activity through Attestcoin and verifies every transaction against the Creditcoin BlockProver precompile. A borrower cannot claim activity they did not produce.

03
Decide

A bounded AI proposes terms. The on-chain Policy independently validates them. Real ERC-20 tokens move from the liquidity pool to the borrower. The agent's reputation updates — and so does its lending authority.

The properties

Decisions, capital, and reputation — all on-chain

MIRA doesn't claim to be trustworthy. It earns the right to manage capital.

Decisions are cryptographically accountable
Every underwriting factor is backed by an Attestcoin inclusion proof. The AI sees evidence, not assertions. The Policy contract independently validates every decision — the LLM cannot approve a loan above the agent's authority or below the rate floor.
Capital is cryptographically accountable
The LiquidityPool holds real ERC-20 tokens. Origination moves them to the borrower; repayment moves them back. Every transaction is a real on-chain transfer — not a counter increment. The pool's available balance is a live read from the contract.
Reputation is cryptographically accountable
The agent's track record lives in an on-chain ledger it cannot tamper with. Every verified repayment raises its score — and its lending authority. A fresh agent can only lend $25; a proven agent (score 650+) earns $100. Five defaults auto-pause the agent.
We don't trust the AI
Five adversarial attacks prove the system rejects malicious LLM output, fake proofs, wrong borrowers, expired evidence, and insufficient liquidity. The headline is literal: MIRA cannot approve a loan its own Policy contract does not independently validate.

The system

Three components, one trust root

A Next.js frontend, a TypeScript worker, and Creditcoin smart contracts. The worker is the only component that holds keys; the frontend never signs Creditcoin transactions, and every factor it shows was verified by the BlockProver precompile. The AI proposes — the Policy disposes.

  • Frontend — borrower flow and agent reputation dashboard.
  • Worker — Attestcoin proofs, bounded AI decisioning, loan lifecycle.
  • Contracts — Policy, Loan, and reputation ledgers on CC3 Testnet.
Read the docs

Borrower wallet

Ethereum Sepolia · real financial activity

Attestcoin proof

MIRA worker

The AI proposes — the Policy disposes

  1. 1  fetch + verify Sepolia activity via Attestcoin
  2. 2  bounded AI proposes loan terms
  3. 3  on-chain Policy independently validates
  4. 4  originate: real ERC-20 tokens move
originate · real token transfer

Decisions

Policy validates

Capital

ERC-20 custody

Reputation

On-chain ledger

repay / default

Reputation → Capital authority

Repayment raises the score → higher lending authority. Default lowers it → less capital. 5 defaults → auto-pause.

feedback loop: the agent earns and loses the right to lend

Creditcoin CC3 Testnet

Policy · Loan · AgentReputation · LiquidityPool

Ethereum Sepolia

verified activity · writability on default

Live

Verification status

The Attestcoin read path is exercised against live CC3 Testnet. The card below reflects the most recent validation run.

Checking validation status…

MIRA