Build AI agents that can handle real capital

Independent verification for AI agents handling real capital. Drop-in escrow, intent checking, and audit trails that take agents from sandbox to production.

How It Works

Agent proposes

Code inputTypescript

await nava.verify({
  intent: 'Swap 5,000 USDC for ETH on Uniswap',
  tx: {
    protocol: 'uniswap',
    function: 'exactInputSingle',
    amountIn: '5000 USDC',
    amountOutMin: '1.47 ETH',
  }
});

Any framework, any model. Your agent reasons, plans and proposes transactions without constraints.

Arbiter verifies

Terminal outputJSON

{
  "decision": "PASS",
  "checks": {
    "intent_alignment":      "PASS  0.98",
    "token_validation":      "PASS  1.0",
    "slippage_tolerance":    "PASS  0.98",
    "sanctions_screening":   "PASS  1.0",
    "adversarial_detection": "PASS  0.97"
  },
  "explanation": "USDC→ETH swap matches intent. Slippage within bounds."
}

A separate model checks the proposal before it goes through.

Escrow executes

(OPTIONAL)

Terminal outputJSON

ESCROW 0x8f3a...c291

Status     HELD → RELEASED
Deposit    5,000 USDC
Verdict    PASS (9/9 checks)
Executed   USDC → 1.47 ETH
tx         0x4b2c...a803

Funds sit in escrow until verification passes. If the Arbiter rejects, nothing moves.

Auditable operations

Terminal outputJSON

Transaction #4,291
Intent: "Swap 5,000 USDC to ETH"

  slippage_tolerance     FAIL  "0.8% exceeds 0.5% limit"
  intent_alignment       PASS
  token_validation       PASS
  sanctions_screening    PASS
  adversarial_detection  PASS

  REJECTED → Agent re-proposed with tighter routing

  Re-verification: 9/9 PASS
  APPROVED  tx: 0x8f3a...c291

Every decision comes with the reasoning behind it, down to which check failed and why. Auditable by compliance, verifiable by users.

Product Stack

Nava SDK

Double-check your agent's work.

Submit transactions to the Nava Arbiter before they execute
Get a signed pass/fail with the reasoning behind it
Works in TypeScript, Python, REST, or as an MCP Server

Get Started Github

Code inputTypescript

import { NavaClient } from '@navalabs/sdk';

const nava = new NavaClient({ apiKey: 'your-key' });

const result = await nava.requestAndAwaitVerification({
  humanIntent: 'Swap 1000 USDC for ETH with max 1% slippage',
  proposedTx: {
    to: '0x3fC91A3afd70395Cd496C647d5a6CC9D4B2b7FAD',
    data: '0x...',
    value: '0',
  },
});

console.log(result.decision);     // APPROVED or REJECTED
console.log(result.explanation);  // why

Execution Escrow SDK

Funds stay put until the Arbiter says go.

Agent proposes, escrow holds, Arbiter gates release
Non-custodial. Your keys, your wallet
For institutions and high-value flows

Integrate escrow Contract reference

Code inputTypescript

import { ExecutionEscrowCore } from '@navalabs/execution-escrow-core';

const escrow = new ExecutionEscrowCore(config);
await escrow.init();

const result = await escrow.requestVerification({
  to: '0x3fC91A3afd70395Cd496C647d5a6CC9D4B2b7FAD',
  value: '0',
  description: 'Swap 1000 USDC for ETH, max 1% slippage',
  escrowAddress: '0x...',
});

// Funds only move if the Arbiter approves
console.log(result.status);    // APPROVED or REJECTED
console.log(result.analysis);  // Arbiter reasoning

Nava Arbiter

A separate model reviews every transaction your agent proposes.

Token validity, sanctions, decimal scaling, gas limits
Intent alignment, adversarial patterns, parameter coherence
Open source and framework-agnostic. Drop into any agent stack

Read the research How it works

Terminal outputJSON

{
  "decision": "PASS",
  "reason": "Transaction matches intent. Slippage within bounds.",
  "results": {
    "token_validation":      { "status": "pass", "confidence": 1.0 },
    "sanctions_screening":   { "status": "pass", "confidence": 1.0 },
    "slippage_check":        { "status": "pass", "confidence": 1.0 },
    "intent_alignment":      { "status": "pass", "confidence": 0.98 },
    "adversarial_detect":    { "status": "pass", "confidence": 0.97 }
  },
  "explanation": {
    "summary": "USDC-WETH swap matches user intent. 0.8% slippage within 1% limit.",
    "key_issues": [],
    "recommendations": []
  }
}

Nava Chain

Network Layer Shared Intelligence

Your agents get smarter as the network grows:

Error patterns discovered by any agent help protect all agents

Verification accuracy improves with every transaction

Built-in audit trail for your peace of mind

View SDK Docs Github

before

Agents stuck in sandbox environments

Manual oversight required for every trade

No deployment confidence without liability risk

Weeks of defensive coding and testing

after nava

Agents handling real capital in production

Ship autonomous agents with safety guarantees

Built-in verification catches errors before they cost money

Fork, modify, deploy - get end-to-end demos running immediately

Audit trails for accountability

View SDK Docs Github

Start Building

Devs and traders

Fork a reference agent or integrate the SDK.

Institutions

Enterprise-grade verification for autonomous agents.