Security & Trust

The agent cannot approve its own work

An agent here can read, draft, enrich and route all day. It cannot send the email, make the payment, or push to production — not because we told it not to, but because the queue will not hand it that item. Your credentials never reach our cloud, so there is nothing on our side to leak.

Human-in-the-loop

The Golden Rule

The AI does everything up to an external side-effect. Then a human approves. Nothing that leaves your walls — an email, a payment, a post, a write to another system — fires without a recorded human decision.

  • Autonomous for reading, drafting, enriching and routing
  • A hard stop at every external side-effect
  • Every approval is logged for audit
  • No prompt-injected instruction can bypass the gate
The approval boundary
Green is autonomous. The wall is the human gate.
The approval boundary Reading, drafting, enriching and routing happen autonomously; a human approval gate stands between them and any external action. AUTONOMOUS Read & understand Draft response Enrich from data Route to queue HUMAN External action fires

Zero-knowledge secrets

The cloud can’t leak what it never held

Credentials live on your machine, encrypted at rest. External work runs on the local agent. Our control plane orchestrates work items — it never sees your plaintext secrets. That’s a guarantee centralized BPM and iPaaS architecturally can’t make.

Local key custody

Secrets are stored in your OS keychain and encrypted at rest on your device — never uploaded.

Edge execution

Anything touching an external system runs on the agent, from your network — not a shared cloud.

Regulated-ready

Because plaintext never centralizes, the model fits PHI and other regulated data by design.


Trust pillars

Verifiable, not just assertable

It runs on your hardware

The agent runs on your hardware, so you can watch what it does: its process, its network calls, its access to your files. Source opens at general availability. Until then, trust the boundary rather than our word for it — the cloud never receives a plaintext credential, so there is nothing there to leak.

Full audit trail

Every move on a work item is recorded: who did it, when, and whether the actor was a person or an agent. You can reconstruct who approved what. What happens inside a model's own reasoning is not recorded — the trail is of actions taken, not thoughts had.

Tenant isolation & RBAC

Multi-tenant organizations keep each team’s data separate, with role-based access controls and scoped caller identity.

Enterprise key recovery

Zero-knowledge by default; on Enterprise, admin-recoverable secrets provide continuity without handing plaintext to the cloud.

Automation you can actually trust.

Human approval on every external action, zero-knowledge secrets, and a client that runs on your hardware rather than ours.