Governance

OpenDoor is the inline proxy. A completion does not go to a provider until the gateway has checked guardrails, the model registry, and your org policies.

That is the product — not a bundle of LiteLLM + OPA + a separate GRC dashboard.

Live path

1. Request — POST /v1/chat/completions with an API key. Optional X-Data-Class: public, internal, confidential, or restricted (default internal).
2. Guardrails — injection blocks the call. PII and secrets are flagged and redacted.
3. Registry — a model marked pending, rejected, or deprecated is held. Data classes the model is not approved for are denied.
4. Policies — org rules match model glob, role, and data class. Actions: allow, deny, require approval, or route to a fallback model.
5. Record — denies and holds write policy_violations. Guardrail hits write guardrail_outcomes. Both show in Violations.

The Playground uses the same path. Set the data class in Parameters to see a policy fire.

Trust Center

Dashboard → Governance is the model inventory. Approving a model here updates model_governance and unblocks that id on the next request.

Policies

Rules stored per organisation. Patterns: * (all models), gpt-*, or deepseek-*|qwen-*. Lower priority number is evaluated first.

Approvals

A pending model returns 403 with human_approval_required until someone approves it. There is no separate workflow engine — the hold is the gateway response.

Violations

Live blocks from the path above. Resolve a row when the team has handled it. The sidebar badge counts unresolved rows.

Compliance

Control mappings and automated reports against GDPR, EU AI Act, ICO UK, and NIST. Evidence and scores live here. They do not replace gateway enforcement.

Sector packs

An industry bundle. Apply writes gateway policies for that data class (human-approval holds and banned-use denies). Recommended models still need Trust Center approval.

Headers