Most enterprise AI can answer anything and touch nothing. Sidekick Machines is the platform that lets agents do real work inside your systems — scoped to exactly what they may reach, approved on the steps that matter, recorded once, and reversible. It installs into your own cloud, single-tenant, and you build on a small, versioned contract designed to outlast the model underneath it.
Governance isn't a policy you bolt on. It's the primitive you build on.
For enterprise platform and security teams — request early access.
One note when the platform opens to design partners. No spam.
Every action an agent takes passes through one boundary. Nothing reaches your systems around it.
Agents carry arguments — never credentials or capabilities. Identity is injected at the edge, reach is deny-by-default, human approval binds to the exact action a person signed, and every step lands on a durable, tamper-evident ledger. Total mediation, not trust.
A deny-by-default boundary every tool call must pass through. Agents carry only arguments; identity is injected at the edge and reach is confined to exactly what's granted. Human approval binds to the precise action that was reviewed and signed — fail-closed by default.
A lifecycle ledger commits each side effect once and only once — across retries, crashes, and failovers. Every request, approval, execution, and denial is written to a durable, tamper-evident trail: who caused it, what ran, where. Actions are designed to be undone.
Installs into your own tenant — your cloud, your data, your keys, never a shared backend. You build as a module author on one small, semantically versioned contract that absorbs SDK and model churn. In-place upgrades stay N-1 compatible, so what you shipped last quarter keeps running.