Admina: Open Source framework for AI governance

Noferi releases Admina, an Open Source framework sponsored by noze for controlling and governing AI model calls, under Apache 2.0 license.

AIOpen SourceCompliancenozeGovernance AdminaOpen SourceAI GovernanceApache 2.0PythonRustEU AI ActProxy

The release

Stefano Noferi releases Admina, an Open Source framework sponsored by noze and distributed under the Apache 2.0 license. Admina is created to address a concrete need: governing the interactions between applications and artificial intelligence models with transparency, control and regulatory compliance.

The architecture

Admina offers two complementary operational modes:

  • Python SDK (in-process): direct integration into application code, intercepting model calls with no network overhead.
  • Transparent network proxy: a proxy that intercepts model calls at the network level, applicable to any application without source code modifications.

Both modes apply bidirectional policies with four actions: ALLOW, BLOCK, REDACT and CIRCUIT BREAK.

The four domains

The framework covers four functional domains:

  • Data Residency: automatic redaction of PII (personally identifiable information) and control of data flows toward external models.
  • AI Infrastructure: native support for Ollama, vLLM, local RAG and on-premise AI infrastructures.
  • Agent Security: protection against prompt injection, automatic loop interruption, agent sandboxing.
  • Compliance: conformity with the EU AI Act, forensic black box for complete audit trail, integration with OpenTelemetry for observability.

The technologies

Admina is developed in Python and Rust — Python for the SDK and integrations, Rust for the high-performance proxy.

The reference website is admina.org.

Need support? Under attack? Service Status
Need support? Under attack? Service Status