Trust Center

AI agents your auditor can also trust.

Every agent action is attributable, contemporaneously logged, hash-chained with HMAC anchored to GCP KMS, approval-gated above customer-defined thresholds, runtime scope-restricted, and exportable as auditor-grade evidence.

Controls overview

The eight load-bearing controls that make every agent's work defensible — all shipped today.

ControlHow it worksStatus
Tamper-evident audit logEvery agent action is appended to a hash-chained audit log signed with an HMAC key rooted in Google Cloud KMS. Any retroactive edit breaks the chain and is detectable by replay.Shipped
Separation-of-duties approvalsSensitive agent changes require a second-party approval. The owner/admin who edited the agent cannot also approve the change — enforced server-side at decision time.Shipped
Per-agent runtime policiesEach agent runs against a declarative policy: allow-listed tools, cost ceilings, network-egress allow-list, and approval gates. Violations short-circuit the tool call before any side effect.Shipped
Evidence exportOwners can export a signed ZIP for any date range — manifest, body.tsv, and HMAC signature — suitable for auditor handoff or regulator request.Shipped
Compliance-profile flipSwitching an org to the Regulated Industry profile raises retention to 7 years, forces PII filtering on, and pins the audit-signing version. The flip is itself an audit event.Shipped
PII redactionPrompts and responses pass through a configurable redactor before reaching upstream model providers. Detected SSNs, account numbers, and free-text PII are replaced with type-preserving placeholders.Shipped
Retention sweepA nightly sweep deletes records past the org’s retention horizon and prunes the audit chain forward, leaving a signed checkpoint. Every deletion is logged with a tombstone entry.Shipped
KMS-rooted signingAll audit signatures derive from a Cloud KMS-managed signing key. Key rotation is supported, and previous signature versions remain verifiable indefinitely against archived public material.Shipped

What we do not claim

Honest > shiny. Where we don't yet have the certification or the property, we don't claim it.

  • ❌ CMMC
  • ❌ FedRAMP
  • ❌ ITAR
  • ❌ GovCloud
  • ❌ SOC 2 (in progress; not certified)
  • ❌ DCAA-certified
  • ❌ "Immutable" (we say tamper-evident)
  • ❌ "Blockchain" (we say hash-chained, append-only)

Stripe & PCI posture

Payments are processed via Stripe Checkout. Cardholder data (PAN) never touches our infrastructure; Stripe Checkout handles the entire card-capture surface. We hold a Stripe customer id and a subscription id, nothing more.

Subprocessors

The third parties that touch customer data in the course of normal operations.

VendorUse
GCPPrimary cloud, KMS, storage, Firestore.
Temporal CloudWorkflow orchestration.
AnthropicClaude family models.
OpenAIGPT family models.
FirebaseAuth + client-side state for the admin app.
StripePayments (Stripe Checkout; no PAN on our infra).
SendGridTransactional and inbound email.
MercuryCustomer banking integrations (read-only via OAuth where used).

Contact

Security reports, compliance questions, and auditor crosswalk requests all reach our security team.

Last reviewed against the shipped product: May 29, 2026