Give Your AI Agent a Wallet. It Can Do the Rest.
Why we built an open-source CLI wallet for AI agents — and how it works
What is it?
AgentsWallets is an open-source CLI wallet tool. Agents install it via npm, create wallets, manage keys, and transact — all locally. No browser. No UI. No cloud dependency.
All private keys are encrypted with AES-256-GCM and stored on the agent's own machine. The security model borrows from the Principal-Agent framework — humans set the rules (Policy Engine), agents operate autonomously within those boundaries.
Built for agent consumption from the first line of code: llms.txt on the website so agents can read what we do, CLI --help as a mini-doc agents call before every command, JSON output by default in non-TTY environments, error codes with recovery_hint so agents can self-heal.
Don't take our word for it — let Claude Code, OpenClaw, or any AI agent try it, evaluate it, and give feedback. They can all pick it up immediately.
Design Philosophy
Agent priorities are the inverse of human priorities:
| Agents | Humans | |
|---|---|---|
| 1 | Determinism | Brand |
| 2 | Atomicity | Docs |
| 3 | Discoverability | Discoverability |
| 4 | Docs | Atomicity |
| 5 | Brand | Determinism |
We inverted the priority stack. Every design decision starts with one question: how does an agent use this?
This is not a traditional wallet with an AI feature bolted on. It was built for agents from Day 1.
Why we built this
Two ideas converged.
Boris (Claude Code lead) @bcherny said: stop making models fight models. Build products for the agent models arriving in 6 months.
Andrej Karpathy @karpathy said: CLI is the natural interface for AI agents.
We combined both: if agents will operate autonomously onchain, they need wallets they can actually install and use. Not a hosted API. Not a browser plugin. A local tool that works even when every server is down.
Existing wallets are all built for humans. MetaMask needs a browser. Phantom needs a UI. Coinbase AgentKit depends on cloud infrastructure. None of them work when your user is an AI agent running in a headless terminal.
What it does today (v0.3.1)
- 6 chains: Ethereum, BNB, Base, Polygon, Arbitrum, Solana
- HD wallet: one mnemonic generates both EVM and Solana addresses
- Policy Engine: transaction limits, token allowlists, spending caps
- Multi-chain balance aggregation in one call, USD pricing
- Polymarket prediction market trading
- AES-256-GCM encrypted local key storage
- Structured JSON output with idempotency keys
- 350+ tests passing, open source, published on npm
Try it
npm install -g @agentswallets/cli
aw init
aw wallet create --name my-agent
aw wallet balance my-agentFour commands. Your agent now has a wallet across 6 chains.
How it was built
One person. AI-native development workflow. Claude + ChatGPT for product design and architecture. Claude Code for implementation. Codex for code review. OpenClaw for product walkthrough, UX testing, and feedback.
The product itself is proof of what AI-era productivity looks like — from idea to production, AI was involved in every step.
What's next
- OKX OnchainOS DEX API integration (swap, bridge, market data)
- Hyperliquid perpetual trading integration
- x402 payment protocol integration
- Wallet-as-a-Service: hosted Policy Engine, audit logs, webhook notifications
- Fiat on-ramp via MoonPay
- AI trading strategy subscriptions