Cortex
The open-source AI gateway that redacts, pseudonymizes and audits every exchange between your people and language models.
Architecture
Cortex gateway
- Authentication
- DLP
- Pseudonymization
- Routing
- Audit
Everything Cortex does
Fail-closed DLP
A versioned registry of built-in and tenant-defined patterns; a blocking pattern refuses the request before any provider.
Advanced detection
Named-entity recognition and local-LLM detection on the optional node.
Pseudonymization
The same value keeps the same tag within a conversation, without letting a database reader correlate it across conversations.
Tamper-evident audit
Hash-chained, append-only at the database level, with a verifier, an auditor console and export.
Tenant isolation
Schema-per-tenant PostgreSQL with row-level security; the tenant is resolved server-side.
Encryption at rest
Envelope encryption with one key per tenant; destroying the key erases the tenant.
OpenAI-compatible
The /v1 routes are a drop-in for OpenAI clients; outbound, OpenAI and Anthropic formats, Ollama and vLLM.
Routing and reliability
Failover chains, retries that never replay a stream in flight, replica load balancing.
Response cache
Exact-match cache on the redacted prompt, per tenant and encrypted; off by default.
Usage and costs
Per-request cost estimate, budgets enforced before sending, virtual keys with rate limits.
Knowledge base
Document collections with per-department access and a manager role.
Enterprise sign-in
OIDC and SAML with your identity provider, local accounts with TOTP MFA.
Retention and erasure
Per-tenant policies with exact preview, scheduled purges, two-step confirmed erasure.
Observability without content
Metrics and structured logs with no conversation content.
French and English
Complete interface in both languages.
Licence
Cortex's core is MIT-licensed: you can use, modify and audit it freely. A commercial version is in preparation.