# Admina

> Admina is the open framework for governed AI development. It ships an
> SDK, an OpenAI-compatible governance gateway (POST /v1/chat/completions,
> streaming and non-streaming, plus GET /v1/models), a transparent proxy,
> a plugin system, a CLI, and a dashboard — delivered as a single install
> with a hybrid Python + Rust engine. Every AI interaction is governed:
> PII redacted, injections blocked, loops broken, actions audited,
> EU AI Act / NIS2 / GDPR compliance tracked. Apache 2.0.
> Pre-1.0 (v0.11.x).

Admina is organised into four governance domains — Data Sovereignty,
AI Infrastructure, Agent Security, Compliance — applied bidirectionally
to every request and response. Governance also covers streamed responses:
deltas pass through a windowed recomposition buffer (StreamRedactor) that
redacts PII spanning delta boundaries, on both the SDK GovernedModel.stream()
and the gateway's SSE surface. Two scores quantify governance posture:
the Admina Score (live runtime composite) and the OISG Adequacy Score
(static capability assessment, Open · Intelligent · Secure · Governed).

## Documentation

- [Documentation overview](https://admina.org/docs/): Card grid linking to every docs section
- [Quick Start](https://admina.org/docs/quickstart/): Install via PyPI (admina-framework) + the three admina dev modes (zero-Docker local, --stack, --with-llm); packaging extras include the new [presidio] extra, with [all] now rolling up [proxy,nlp,telemetry,adapters,presidio]
- [Configuration](https://admina.org/docs/configuration/): admina.yaml options, env vars, encrypted vault, three forensic backends (memory / filesystem / s3; MinIO reachable via the s3 backend); v0.11.0 adds ADMINA_PII_ENGINE, ADMINA_GUARD_FAIL_MODE, ADMINA_FORENSIC_STATE_KEY (HMAC-signed forensic chain state) and ADMINA_GATEWAY_UPSTREAM / ADMINA_GATEWAY_BLOCK_MESSAGE / ADMINA_GATEWAY_MODELS_ALLOWLIST
- [Governance Domains](https://admina.org/docs/domains/): Data Sovereignty (PII redaction with a selectable engine — default spacy-regex, optional Microsoft Presidio via ADMINA_PII_ENGINE=presidio, EN + IT — plus residency zones), AI Infrastructure (LLM engine, RAG, WebUI), Agent Security (33-pattern injection firewall across 9 categories, loop breaker), Compliance (EU AI Act + NIS2 + GDPR + cross-regulation matrix, post-Omnibus VII timeline)
- [API Reference](https://admina.org/docs/api/): REST + WebSocket + Prometheus /metrics, including the dashboard endpoints, multi-regulation compliance endpoints, the /mcp JSON-RPC governance proxy, the OpenAI-compatible governance gateway (POST /v1/chat/completions streaming + non-streaming, GET /v1/models) and GET /api/v1/forensic/verify; /api/v1/validate returns action = ALLOW | BLOCK | REDACT (REDACT replaced MODIFY in v0.11.0, breaking, no shim)
- [Integrations](https://admina.org/docs/integrations/): framework integrations (LangChain, CrewAI, n8n, OpenClaw, Cheshire Cat AI, GuardrailsAI) + model provider adapters (OpenAI, Ollama, Anthropic, Mistral, AWS Bedrock, Google Gemini, vLLM — since v0.10.0, per-provider extras); BaseModelAdapter.send_stream() since v0.11.0 — real streaming on OpenAI, Ollama, vLLM and Anthropic, single-chunk base fallback on Mistral, Gemini and Bedrock
- [Rust Engine](https://admina.org/docs/rust-engine/): opt-in admina-core (single abi3-py311 wheel), ADMINA_ENGINE=auto|python|rust selector, build from source, benchmarks (6.25µs full-pipeline median); Rust and Python engines are not equivalent (Python has broader detection coverage) and the gap is measured, not asserted — a shipped red-team detection-efficacy suite (admina/redteam, runnable via scripts/redteam.py) scores injection, PII and loop detection against committed corpora with a pinned baseline and a CI regression gate

## Metrics

- [Admina Score](https://admina.org/docs/admina-score/): Live-runtime composite (0–100) — formula, the five components, how to improve each
- [OISG Adequacy](https://admina.org/docs/oisg/): Open · Intelligent · Secure · Governed — 4 pillars × 5 criteria; the full Docker Compose stack ships OISG-adequate (85), a bare pip install scores 75 (Good coverage)
- [Admina Score vs OISG Adequacy](https://admina.org/docs/admina-score-vs-oisg/): Live runtime vs static capability — comparison, 2×2 interpretation matrix, scenario-based guidance

## Project

- [Community](https://admina.org/community/): Roadmap (0.9.x–0.11.1 shipped → 0.12.0 observability & performance → 0.13.0 multi-tenancy/RBAC → 0.14.0 compliance templates → 1.0.0 LTS), how to contribute, maintainer
- [Sponsor](https://admina.org/sponsor/): GitHub Sponsors (individuals) + enterprise sponsorship channel — funds core development, security audits, EU AI Act compliance research, project infrastructure
- [Security Policy](https://admina.org/security/): Coordinated disclosure (info@admina.org), 48h ack, 5 biz-day assessment, 10 biz-day fix timeline, supports the latest minor line (v0.11.x, pre-1.0)
- [Heimdall](https://admina.org/mascot/): The owl mascot of Admina

## Source

- [GitHub repository](https://github.com/admina-org/admina): Apache 2.0, full source code
- [PyPI · admina-framework](https://pypi.org/project/admina-framework/): SDK + proxy distribution (current: 0.11.1, released 2026-08-07)
- [PyPI · admina-core](https://pypi.org/project/admina-core/): Optional Rust accelerator (abi3-py311 wheel, version-synced with admina-framework)
- [CHANGELOG](https://github.com/admina-org/admina/blob/main/CHANGELOG.md): Per-version release notes
- [Releases](https://github.com/admina-org/admina/releases): Tagged releases on GitHub (latest: v0.11.1)
- [ROADMAP](https://github.com/admina-org/admina/blob/main/ROADMAP.md): Forward plan toward 1.0.0
- [MODEL_CARD](https://github.com/admina-org/admina/blob/main/MODEL_CARD.md): Transparency document — scope, limitations, known failure modes of each Admina component
- [SECURITY](https://github.com/admina-org/admina/blob/main/SECURITY.md): Repo security policy

## Optional

- [oisg.ai](https://oisg.ai): External OISG paradigm reference
- [EU AI Act Omnibus VII press release](https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/): Council/Parliament agreement of 7 May 2026 — postponed deadlines reflected in Admina's EU_AI_ACT_DEADLINES table