IntentPay

IntentPay is an AI-assisted Web3 transaction prototype. It translates natural-language payment intent into a structured transaction plan, then keeps execution behind an explicit wallet review step.

The project explores a safer product boundary for AI and blockchain applications: the model can interpret and prepare actions, but the user must still review and approve any irreversible on-chain transaction.

• Status: Prototype
• Portfolio case study: https://subhajitpradhan.vercel.app/projects/intentpay
• Inspect the implementation: ai-agent.js, wallet-manager.js, and frontend/

Why This Exists

Crypto wallets expose powerful actions through low-level addresses, gas settings, contract calls, and transaction prompts. That is difficult for non-technical users and risky for teams that need guardrails.

IntentPay experiments with a higher-level workflow:

1. A user describes what they want to do in natural language.
2. The backend resolves intent, contacts, balances, and guardrails.
3. The system prepares a structured transaction preview.
4. The wallet remains the final execution boundary.

Core Flows

• Natural-language payments such as split payments and named recipients.
• Handle and ENS-style recipient resolution.
• Transaction preview before wallet execution.
• Gas estimation and structured transaction errors.
• Spending guardrails and whitelisted recipient concepts.
• Transaction history with status and hashes.
• Security event logging for suspicious or failed actions.

Architecture

React client
  Intent input, transaction preview, wallet-facing UI

Express API
  Auth, contacts, guardrails, transaction planning, sanitized responses

LLM adapter
  Converts user intent into structured tool calls and transaction plans

Wallet/blockchain layer
  Ethers.js, Sepolia RPC, gas estimation, transaction submission boundary

Security layer
  AES-256-GCM encrypted wallet data, bcrypt password hashing, event logs

Safety Boundary

IntentPay intentionally separates AI planning from blockchain execution.

• The AI does not silently execute transactions.
• The API sanitizes responses so private keys and password hashes do not leave backend logic.
• Transaction details are prepared for review before signing.
• Guardrails can restrict recipients and spending limits.
• Failed or suspicious flows are recorded through security events.

Tech Stack

• Node.js
• Express.js
• React
• Vite
• Ethers.js v6
• OpenRouter-compatible LLM API
• AES-256-GCM
• bcrypt

Getting Started

Prerequisites

• Node.js 18+
• OpenRouter API key
• Sepolia RPC endpoint
• Sepolia test ETH

Install

npm install
cd frontend
npm install

Environment

Create a .env file in the repository root:

PORT=3001
RPC_URL=https://ethereum-sepolia-rpc.publicnode.com
ENCRYPTION_KEY=your_32_char_random_key_here
OPENROUTER_API_KEY=your_openrouter_key_here
LLM_MODEL=meta-llama/llama-3.1-8b-instruct:free

Use a dedicated test key and testnet-only funds when experimenting locally.

Run

node server.js

In a second terminal:

cd frontend
npm run dev

Open:

http://localhost:5173

Example Prompts

• "Remember @alice is 0x742..."
• "Send 0.01 ETH to @alice and 0.02 ETH to vitalik.eth."
• "Set my daily limit to 0.5 ETH."
• "What is my security status?"
• "Show my transaction history."

Tradeoffs

• Server-side wallet storage is useful for prototyping custodial flows, but production systems should use HSM, MPC, account abstraction, or non-custodial signing models.
• Natural-language UX reduces friction, but every generated action still needs visible review because blockchain transactions are irreversible.
• LLM output must be treated as untrusted input and validated before it reaches transaction-building code.

Roadmap

• Move sensitive key handling to HSM, KMS, MPC, or smart-account infrastructure.
• Add ERC-4337 account abstraction for recovery, policies, and gas sponsorship.
• Expand guardrails with time locks, approval thresholds, and multi-sig policies.
• Add multi-chain support after the safety model is stable.
• Add tests around intent parsing, guardrail enforcement, and transaction validation.