Admina: Open Source framework for AI governance

Admina v0.9.1 released (20 May 2026): Apache 2.0 Open Source framework for AI governance, created by Stefano Noferi and sponsored by noze. Python SDK + Rust proxy, 4 governance domains, ~6µs overhead, 700+ tests.

AIOpen SourceCompliancenozeGovernanceAdminaOpen SourceAI GovernanceApache 2.0PythonRustEU AI ActProxyOISGMCP
Diagram of the Admina governance pipeline: the request leaves the application and passes through the anti-injection firewall, the PII redactor and the loop breaker before reaching MCP Server or LLM, with BLOCK, REDACT, CIRCUIT BREAK and ALLOW outcomes and a SHA-256 hash chain recording every decision
Every request goes through the three scanners in sequence before it reaches the model, and the response runs back along the same chain. The latencies hold with the Rust engine loaded: on the pure Python fallback the same pipeline measures around 2.3 milliseconds. Diagram based on docs/guides/architecture.md from Admina v0.9.1.

The release

On 20 May 2026, Admina v0.9.1 (pre-1.0) is released — an Open Source framework for AI governance created by Stefano Noferi and sponsored by noze. The project is distributed under the Apache 2.0 license, published on github.com/admina-org/admina, with official site at admina.org.

The stated goal is to enable building AI applications that are governed by design: a single install bundles SDK, proxy, CLI, dashboard and a plugin system to apply bidirectional governance on every interaction between the application and the models.

The architecture

Admina runs in two complementary modes:

  • Python SDK (in-process): imported as a library via GovernedModel, intercepting model calls directly in application code — no network overhead.
  • Network proxy (Rust + Python): a transparent proxy applicable to any client with no code changes; it talks to MCP Server, LLM API (Ollama, vLLM), ChromaDB and MinIO.

Both modes apply bidirectional policies — on both request and response — with four actions: ALLOW, BLOCK, REDACT, CIRCUIT BREAK. Supported sources include the SDK, generic AI agents, OpenClaw and frameworks like LangChain and CrewAI.

The four governance domains

  1. Data Sovereignty — PII redaction, data residency enforcement, data classification. Rust-accelerated at 0.62µs per scan.
  2. AI Infrastructure — LLM engine (Ollama, vLLM), RAG pipeline (ChromaDB), web UI, hot model switching. Opt-in.
  3. Agent Security — anti-injection firewall (15+ patterns, 2.08µs), loop breaker via TF-IDF cosine similarity (2.38µs). Bidirectional.
  4. ComplianceEU AI Act classification (Art. 6-15), forensic black box with SHA-256 hash chain on MinIO, native OpenTelemetry integration.

Performance

Typical governance overhead across the full 4-domain pipeline: median 6.25µs, P95 7.04µs, P99 7.29µs (Rust engine, Python fallback). The v0.9.1 release ships with 700+ tests in the automated suite.

OISG adequate by default

Admina is built around the OISG paradigm — Open, Intelligent, Secure, Governed — published on oisg.ai. The adequacy score is computed at runtime from the live installation state, not from static checklists: a default install reaches the adequate band.

Technology stack

The core is written in Python (SDK, integrations) and Rust (proxy, scanners, firewall). The local dev stack includes dashboard (:3000), proxy (:8080) and Grafana (:3001). Distribution details (package and CLI commands) will be announced at release.


Stefano Noferi - Founder e CEO/CTO di noze
Tech Entrepreneur - AI Governance & Security Architect

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