Explicit principals
Humans, agents, services, and system jobs are distinct identities with distinct ownership and grants.
Security architecture
Architecture your security team can interrogate.
Onterix is being designed around explicit principals, same-or-stricter source access, data-plane credential isolation, destination-aware protection, exact-action approvals, and reconstructable evidence.
WHOActor · subject · agent · clientidentity
WHATCapability · source · resource · payloadscope
WHERERuntime · region · model · recipientdestination
WHYPurpose · policy · approval · reasondecision
PROOFVersions · hashes · refs · outcomeevidence
Security invariants
Controls that do not become paid bypasses.
Baseline tenant isolation, permission checks, approval enforcement, redaction, idempotency, and audit integrity are intended as platform security foundations—not optional shortcuts.
Intersection, never union
On-behalf-of access intersects the subject, actor, client, credential, source, and tenant policy.
Source-native ceiling
When a source can express permissions, the runtime cannot manufacture broader authority.
Credential isolation
Agent hosts and model providers never receive downstream connector credentials.
Exact-action approval
Approval authorizes one eligible, versioned payload and cannot expand the execution boundary.
Safe evidence
Audit records safe summaries, hashes, versions, findings, and refs—not blind copies of sensitive payloads.
Evidence chain
Reconstruct the decision—not just the API call.
A useful audit record connects identity, source permissions, client trust, protection findings, policy, transformations, approval, execution, and outcome without leaking the sensitive values it is meant to protect.
01Request and identity boundintent
02Source and destination resolvedcontext
03Protection providers evaluatedfindings
04Policy and approval recordeddecision
05Execution and result correlatedoutcome
Customer-controlled plane
Keep sensitive state out of the shared service.
Strict deployment modes are intended to keep the content-bearing execution path and its sensitive operational state in the customer environment.
- Connector credentials and customer database secrets
- Source content, indexes, chunks, and embeddings
- Protection findings, token vaults, and quarantined artifacts
- Workflow payloads, approval evidence, and execution-local state
- Customer model routes, KMS/HSM bindings, and audit sinks
- Time-bound, tenant-approved support access only
Trust status
Private preview means saying exactly what is ready.
The architecture and deployment modes on this site are target commitments under active implementation. Onterix does not claim certifications, service levels, production availability, or verified client and connector coverage that have not been completed.
| Area | Current public status | How we present it |
|---|---|---|
| Runtime | Runnable Phase 0 scaffold and fixtures | Private preview development |
| Connectors and clients | Readiness varies; reference and target paths | Labeled per surface |
| Deployment modes | Target architecture; not generally available | Planned |
| Compliance certifications | No public certification claim | Not claimed |
| Production SLA and recovery | Internal targets, not a customer SLA | Not claimed |
Security review
Interrogate the design before trusting the runtime.
Design partners receive a scoped architecture and security review tied to the workflow, connectors, data classes, providers, and deployment boundary being evaluated.