Agents that can only send what a person can read

An open-source TypeScript toolkit and MCP server, in development. It turns the ERC-7730 descriptors wallets use into tools an AI agent can call, and puts every write through simulate, decode, clear-sign and confirm before anything is signed.

Tell us what you would use it for

In development

Every write takes four steps

Simulate. The call runs against current chain state first, so the agent and the person see its effects before they are real.
Decode. The calldata is decoded against the contract’s verified ABI.
Clear-sign. The contract’s ERC-7730 descriptor turns the decoded call into the sentence and labeled values a wallet would show.
Confirm. A person reads that text and approves this one transaction. Without that approval, nothing is signed.

What it is for

Protocols that want a descriptor drafted from their ABI and Sourcify metadata, and checked with the official ERC-7730 tooling, in their own CI.
Agent builders who want a typed tool for each described function, so an agent can only propose calls a descriptor explains, served over MCP with the four steps built in.

Built on what exists

The toolkit wraps the tooling the ecosystem already maintains, such as Ledger’s python-erc7730, the Clear Signing libraries and Sourcify’s APIs, rather than replacing it, and it uses viem for chain access instead of competing with it.

Want to shape it before it ships? Tell us what you would use it for.