# Blockchain infrastructure for AI agents

An agent that can read a chain is useful. One that can also move value is a
different thing entirely — and the reason most teams stop at read-only is that
handing a private key to a model is not an option.

It does not have to be. The agent prepares the transaction and broadcasts it;
the signature happens with a key it never sees.

## What an agent can do

- **Read the chain** — balances, transaction history, blocks, token holdings and
  market rates across 35+ networks, through MCP tools, a CLI, or plain REST.
  See [AI Compatibility](https://cryptoapis.io/products/ai-compatibility).
- **Act on it** — prepare an unsigned transaction, sign it with a key the agent
  never holds, then broadcast. See
  [Offline Signer](https://cryptoapis.io/products/offline-signer).
- **Pay for what it uses** — settle for an API call in USDC over the open x402
  protocol, on-chain and non-custodial. See
  [x402 Payments](https://cryptoapis.io/products/x402).
- **See it working** — the guide runs end to end on the free tier. See
  [Use Crypto APIs with AI](https://cryptoapis.io/guides/use-cryptoapis-with-ai).

## The agent never holds your keys

Addresses are derived from an xPub — a public key — so nothing the agent can
reach will move funds. Transactions are prepared by the API and signed by
`@cryptoapis-io/offline-signer` inside your own process: MIT-licensed, no API
key, no network call. The agent hands the signed result back for broadcast and
never sees the key that produced it.

## Start

The MCP server, the CLI and the whole read surface work on the
[free tier](https://cryptoapis.io/free-crypto-api). No card to start.