Skip to main content
Governance is the layer every request inherits. You set identity, policy, and security rules at the top of your organization, and they cascade down to every department, app, agent, and key — automatically, with no extra code. Set the rules once; they follow every request.

​
Cascading policy

Policies flow through a hierarchy: 
Organization → Department → App → Key
Each level can only tighten, never loosen. A department can’t override an org-level block, and a key can’t grant itself more than its app allows. Set a budget cap or a model allowlist at the org and you know it holds everywhere beneath it.
LevelExample rule
Orgbudget · max $50k / mo
Deptmodel · only gpt-5
Apprate · 60 RPM
Keyscope · read-only

​
Policy types

  • Rate limits — requests per minute (RPM) and tokens per minute (TPM), per key, model, or endpoint
  • Budgets — daily, weekly, and monthly spend caps with automatic enforcement
  • Model allowlisting — restrict which models a key can reach
  • Endpoint restrictions — control which API endpoints a key can call
  • Guardrails — content policies for intent blocking, PII redaction, and toxicity thresholds
  • Custom rules — regex and keyword patterns with block, flag, or reclassify actions
  • Intent rules — per-intent actions such as allow, block, or force optimization

​
KYT (Know Your Tokens)

KYT is DeepMyst’s inline security engine. Every request is classified in under 15ms before it reaches a model, so threats are caught in the request path — not after the fact.

Prompt injection

Detects attempts to override system prompts or exfiltrate instructions.

PII detection & redaction

Identifies personal information and redacts it before forwarding — replace, mask, tokenize, or hash.

Toxicity scoring

Rates content against your thresholds and flags or blocks accordingly.

Intent classification

Categorizes every request so policy can act on what it’s actually trying to do.

​
Two passes

  • Pass 1 (inline, <15ms) runs synchronously in the request path to classify and enforce guardrails before the model is called.
  • Pass 2 (async) runs after the response for deeper enrichment, behavioral analytics, and anomaly detection across requests.

​
Guardrail enforcement

Based on classification, the gateway can:
  • Block the request and return an error
  • Flag it for manual review
  • Redact PII before forwarding
  • Allow it with full logging
Flagged requests land in a review queue where they can be dismissed or escalated.

​
Identity and access

  • SSO — SAML 2.0 and OIDC for enterprise sign-in
  • SCIM — automatic user provisioning and de-provisioning
  • RBAC — role-based access control across owner, admin, operator, team lead, auditor, member, and viewer roles, plus a read-only flag
  • Access scopes — row-level grants that give a team lead read access to specific projects, departments, or users

​
Audit and replay

Every request and every agent run is written to an immutable audit log:
  • Audit-logged forever — inputs, outputs, model, tokens, cost, and policy decisions
  • Replayable — replay any run on demand to reproduce what happened
  • Evals on every output — each result is scored against your rubric

​
Credits and metering

Usage is metered per request and billed in credits (1 credit = $0.01). Every response carries DeepMyst headers — request ID, token counts, compression ratio, cost, credits, rate-limit and budget remaining, classification category, and whether the request was flagged or redacted — so cost and policy are observable on every call.

​
Next steps

The OS Stack

See how governance sits beneath every agent and request

Agent Builder

Build agents that inherit this governance automatically