Cognivern is a spend-governance control plane for AI agent wallets. Every spend is policy-checked before signing—approved, held, or denied based on budget, vendor, chain, and risk—with FHE-encrypted on-chain budgets keeping limits private. A Gemini 3.1 agent runtime follows a PLAN → EVIDENCE → PREVIEW → CONFIRM → EXECUTE → AUDIT workflow with human approval before spending. Deployed on Arbitrum Sepolia and Robinhood Chain Testnet with Hyperlane cross-chain execution and audit evidence anchored to Filecoin and 0G.
AI agents are beginning to spend real money autonomously—paying for API credits, executing trades, procuring services, and interacting with DeFi protocols. Yet teams have almost no visibility or control over that spending. Today, most simply hand an agent a wallet key and hope for the best. Budgets exist as plaintext configuration, there is no enforceable approval workflow, and no regulator or auditor would accept the resulting audit trail. When something goes wrong, teams typically discover it only after funds have already been spent.
Cognivern is a spend-governance control plane for AI agent wallets. Every spending request passes through an on-chain policy evaluation layer before any transaction can be signed. Policies can deny, hold, or approve requests based on budget, vendor, blockchain, and risk. The core innovation is confidential policy enforcement: budgets and spend counters are encrypted on-chain using Fully Homomorphic Encryption (FHE) with Fhenix/CoFHE on Arbitrum Sepolia, meaning spending limits remain private even from anyone with direct RPC access. Every approval, denial, and execution becomes a verifiable on-chain event rather than a backend assertion.
The platform includes a Gemini 3.1-powered agent runtime that follows a deterministic governance workflow:
PLAN → EVIDENCE → PREVIEW → CONFIRM → EXECUTE → AUDIT
Before any real funds move, the agent gathers supporting evidence—including vendor reputation (MongoDB MCP), smart contract audit signals (ChainGPT and 0G Compute), and applicable policy constraints—then presents a complete preview for human approval. Every completed action is anchored to both Filecoin and 0G, creating a tamper-resistant, non-repudiable compliance trail.
The system is deployed across multiple networks, including Arbitrum Sepolia (FHE ConfidentialSpendPolicy, GovernanceContract, GovernedVault) and Robinhood Chain Testnet (GovernanceContract and GovernedVault), with cross-chain execution coordinated through Hyperlane.
Current implementation:
5 Solidity smart contracts
852 Git commits
196 TypeScript files
31+ Vitest test suites
21 live API integration tests
The core Cognivern platform was developed over several months prior to the buildathon, with 852 Git commits, 196 TypeScript files, 5 Solidity contracts, a live production frontend/API, and the FHE-powered ConfidentialSpendPolicy contract already deployed on Arbitrum Sepolia. During the buildathon, the focus was on Arbitrum ecosystem integration, product refinement, and production hardening. GovernanceContract and GovernedVault were deployed and source-verified on Arbitrum Sepolia (Chain 421614) and Robinhood Chain Testnet (Chain 46630), with Robinhood support added across the frontend wallet configuration and UI. The product experience was refined through a problem-first landing page narrative, clearer governance workflows, improved information architecture, stronger visual hierarchy for approvals and audit evidence, and overall UX improvements to make AI spend governance more intuitive and accessible. The production API was hardened for reliability under concurrent usage by adding centralized async error handling, eliminating unhandled promise rejection crashes, and introducing retry logic across backend tests, improving the live API test suite from 12/21 to 17/21 passing. Additional buildathon work included preparing the Arbitrum submission documentation, Product Canvas, and GTM Canvas using the Open House London framework. The positioning focused on Cognivern’s wedge: an AI spend-governance layer with FHE-encrypted budgets, why now: the convergence of AI agents, FHE infrastructure, and regulatory demand, and why on-chain execution is essential for verifiable and confidential governance. The buildathon contribution was not the initial creation of the platform’s core technology, but the deployment, refinement, and ecosystem positioning of Cognivern for Arbitrum through verified infrastructure, improved product experience, production hardening, and clearer market positioning.
n/a
