Silroot is an embedded secure-signing module that lets autonomous agents and machines transact on their own — within limits enforced in silicon. The host can ask. Only your policy, and you, can answer.
Every action an agent proposes is evaluated against your policy inside the secure element — before anything is signed. The host running the agent is never trusted.
Within your caps, rate limits and allowlist, the module signs instantly. No human, no latency — the point of autonomy, preserved.
Over a threshold, the request is held and pushed to your phone — not through the agent's machine. You approve the exact request, cryptographically bound.
Denylisted addresses, disabled skills, or a tripped kill switch die on the device. A prompt-injected transfer never even reaches you.
The signing key lives in the secure element and never leaves it. A separate owner key on your phone holds authority, and it's the only thing you back up.
Generated on-chip, non-extractable. No export path, no seed phrase for the module key. The signing boundary holds even if the host is fully compromised.
Per-transaction caps, daily limits, allowlists, velocity, time windows and per-skill permissions — evaluated in the element. The host cannot override them.
Escalations are signed over the request hash and verified against your registered owner key. A compromised host can't fake an approval or swap the request.
The module is a revocable session key on a smart account. Wipe it and funds are untouched — re-enroll a replacement from your phone in seconds.
Drop in the SDK or expose Silroot as an MCP server and any agent discovers it automatically. Same surface whether you're on Python, TypeScript, or bare firmware.
# the agent never touches a key — it asks the module from silroot import Silroot sig = Silroot.connect() # USB / BLE / MCP result = sig.request_payment( to="api.vendor.eth", amount=5_000_000, # 5 USDC ) if result.status == "escalated": # held on-device, pushed to the owner's phone result = sig.await_approval(result.id) # >>> signed · off-allowlist & injections rejected in silicon
Every action waits for a person to tap a screen. Safe — but it breaks the moment there's no human present, which is the entire premise of autonomous machines.
The machine acts freely inside hardware-enforced bounds and only interrupts you for the exceptions. You supervise the boundary — not every click.
We're shipping the first modules to a small group of builders. Tell us what your machines need to do safely.