Skip to main content
Intent flows through six layers between an LLM tool call and the sensitive data it ultimately reaches. One missing layer is a bypass waiting to happen.

​
L1 — LLM Surface

Tool registry + system-prompt enforcement. The LLM only knows about tools the gateway exposes; system prompts that try to circumvent intention declarations fail at the gateway.

​
L2 — Fastify Gateway

Intercepts every LLM tool call before it leaves your network. The gateway is the choke point — every other layer assumes the gateway has already extracted intent.

​
L3 — Intent Intelligence

A FastAPI extractor that parses the tool call into a structured intention record. Per-user history is keyed by user_id so anomalies surface (“alice never reads SSN; this call wants SSN — flag”).

​
L4 — Policy Layer

Declarative rules that match intentions to allowed elements. Policy always wins over history inference. An intention explicitly allowed in policy is allowed even if it’s anomalous; an intention explicitly denied is denied even if it’s been used before.

​
L5 — Token + Credential Layer

Intent tokens minted from approved intentions. Optional Okta token hooks enrich IdP-issued tokens with the grant claim. Vault issues just-in-time database credentials scoped to the active grant.

​
L6 — DB Proxy

Column-level enforcement at the database. Queries are rewritten or blocked based on the active grant.elements list. Blocked columns are highlighted in the audit log with remediation suggestions.

​
Visual

​
Adoption order

You don’t install all six on day one. The maturity model walks you up the stack — start at L1+L2 (observe), add L3+L4 when patterns are clear, layer L5+L6 when compliance demands enforcement.