Ledger's new open-source kit lets AI agents prepare money moves, with a human still holding the signing key.

An AI agent drafts a payment, prepares a swap, or queues a wallet transfer. The one thing it cannot do is finish the job. A person, on a hardware device, still has to press the button. Ledger, the French hardware-wallet maker best known for the USB sticks that store crypto private keys, is now publishing the plumbing for that pattern as open-source code.

The release, called the Ledger Agent Stack, is the first major delivery in Ledger's stated AI roadmap. It targets developers rather than consumers, and it lands alongside a Ledger-hosted builder event called Ledger N3XT: Agent Stack Build & Show. The pitch is the design. Agents can read balances, build transactions, and propose what should happen. The private key never leaves a Ledger device, and any sensitive action waits for a human to sign on the hardware signer. Ledger calls this "separation of suggestion from execution."

The stack comes in three pieces, all documented under Ledger's developer docs. A Wallet CLI gives an agent programmatic access to send, receive, swap, and track on a Ledger-secured account. Enterprise and Enterprise Multisig CLIs extend the same model to treasury and governance flows, with the signing step bound to hardware security modules instead of consumer devices. On top of that sits the Device Management Kit, or DMK, a set of markdown-based "skills" that let an LLM runtime wire an agent to a Ledger device the way a developer would. The wallet-cli code itself lives inside Ledger's main ledger-live monorepo, where the broader product is maintained.

Two stats from Ledger set up the problem the stack is trying to solve. According to the company's Agent Stack announcement, roughly 60% of security breaches involve some form of human error, and roughly a quarter of publicly available AI agent skills contain at least one security vulnerability. Both numbers appear in the announcement without primary citation, so they read as Ledger's framing of the problem rather than as settled industry data. The bet underneath is that the missing piece in agent-driven finance is not a smarter model but a more honest signing step, with the hardware device as the seam.

The pattern travels beyond crypto. A developer building any agent that touches money, from invoice automation to treasury rebalancing to marketplace payouts, can keep the agent in charge of research, drafting, and queuing, and route the actual transfer through a hardware-rooted approval step. The agent never holds the key. The human, when they choose, signs. The toolkit makes that shape repeatable instead of bespoke.

Ledger is also signaling where this goes next. The Agent Stack is positioned as the foundation for further work on agent identity, agent intents and policies, and what the company calls "proof of human." None of those are shipping yet, and the public repo for the agent-skills project is small, with a few dozen commits and a handful of stars at time of writing. The pattern is adoptable now. The surrounding product is still being built.