The biggest unsolved problem in crypto AI is not intelligence, it is trust.

Most AI agents that interact with blockchain wallets today do so by storing a user’s private key directly in the system.

CoinFello, an AI agent built for smart contract interaction, thinks it has a better way.

On Wednesday, the company launched its open-source OpenClaw skill in partnership with MetaMask, a framework that lets AI agents execute onchain transactions without ever accessing a user’s private key.

Rethinking AI wallet security

The core of the OpenClaw skill is a concept called delegated permissions.

Instead of giving an AI agent broad access to a wallet, users can grant it a narrow, task-specific authorisation, enough to complete one defined action, nothing more.

The system runs on two Ethereum standards: ERC-4337 smart accounts and ERC-7710 delegations, both part of the MetaMask Smart Accounts Kit.

In plain terms, these are technical frameworks that allow a wallet to issue limited, programmable instructions to another party, in this case, an AI agent, while keeping the actual signing key on the user’s own device.

The agent can act, but it cannot take anything it was not explicitly authorised to touch.

CoinFello’s own agents, called Moltbots, act as the user-side layer.

A Moltbot receives a natural-language instruction and passes it to CoinFello as a delegated task.

Before anything executes, an evaluation layer reviews the transaction for validity. Only then does it go onchain.

“If we want agents to participate meaningfully in the onchain economy, we need a security model that is better than handing an autonomous system a private key,” said Brett Cleary, CTO at CoinFello.

“The CoinFello Skill introduces hardware-isolated keys and fine-grained delegations giving AI agents a secure way to execute transactions while helping bootstrap onchain capabilities for the broader agent ecosystem.”

What agents can actually do?

The capabilities are broad.

Using natural-language prompts, Moltbots running the OpenClaw skill can swap ERC-20 tokens, the most common type of crypto asset, bridge funds across different blockchain networks, interact with NFTs, stake assets, lend, automatically rebalance portfolios, and execute multi-step trading strategies.

That range matters as it positions CoinFello not as a single-function tool but as a programmable financial agent capable of managing an active onchain portfolio, all without requiring technical knowledge from the user.

The skill is built on the Agent Skills specification and is compatible with Claude Code environments.

It is released under the MIT licence, meaning developers can freely modify, deploy, and build on it.

A growing ecosystem

The timing reflects real momentum in the OpenClaw developer ecosystem.

In the past two months alone, the OpenClaw GitHub repository has surpassed 150,000 stars and 22,000 forks, strong signals of developer adoption. npm downloads exceeded 416,000 in the past 30 days, a metric that tracks how often developers are actively pulling the code into their own projects.

The broader trend is clear: developers are increasingly experimenting with autonomous software agents capable of interacting with decentralised networks.

What has been missing is a security architecture credible enough for real financial use.

CoinFello’s OpenClaw skill is a direct attempt to build that foundation, not by limiting what AI agents can do, but by defining precisely how far their reach should go.

The post CoinFello launches AI tool to execute crypto trades without private keys appeared first on Invezz