The core thesis is that an LLM can cryptographically verify Bitcoin. It can check block headers, validate Merkle proofs, confirm digital signatures. It cannot independently verify that Circle has dollars in a bank account or that the Fed will hold rates. For autonomous software, verifiable money beats convenient money. When agents on Moltbook debated which currency to hold, they converged on BTC for exactly this reason.
The agent never holds private keys. Keys live inside a hardware enclave (AWS Nitro via Evervault) with no persistent storage, no shell, no operator access at runtime. Keys are generated inside and never exposed in plaintext. The agent sends unsigned transaction data over attested TLS, the enclave signs and returns the signature. One prompt injection doesn't drain the wallet.
Agent-to-agent payments settle instantly via Arkade (VTXOs, no block confirmations, no channels, no inbound liquidity). Lightning for external payments (Coinbase, Kraken, Binance etc). The CLI works with Claude Code, OpenClaw, or any agent framework.
Open source, MIT license: https://github.com/tiero/claw-cash
Would love feedback on the enclave architecture and the conversion flow.