On 20 May 2026 Admina 0.9.1 is published, the first public release of the framework. The release is pre-1.0: the public API is feature-complete and in production use, but may still evolve in response to early-adopter feedback before the 1.0 stability commitment. Licence: Apache 2.0. Repository: github.com/admina-org/admina.

What it is

Admina is a framework for building governed AI applications. A single install ships five coordinated components: an SDK, a transparent proxy, a plugin system, a CLI and a dashboard. Each component applies the same governance rules, both in-process (via SDK) and over the network (via proxy). Integration with existing applications requires no code changes in proxy mode.

Architecture — four governance domains

Rules are organised in four domains applied bidirectionally (on outbound requests and inbound responses):

  • Agent Security — anti-injection firewall (15-regex fast path + heuristic scoring) and loop breaker (TF-IDF cosine similarity)
  • Data Sovereignty — PII redaction (spaCy NER + regex for email, phone, SSN, credit card, IBAN, IP), residency-zone enforcement, data classification
  • Compliance — EU AI Act risk classification (Article 6), gap analysis against Articles 9–15, forensic black box (SHA-256 hash chain with MinIO), OpenTelemetry native spans
  • AI Infrastructure — opt-in: LLMEngine (Ollama / vLLM with GPU auto-detection), RAGPipeline (ChromaDB / Milvus with configurable chunking), WebUI (Open WebUI container with built-in, OIDC or LDAP auth)

Hybrid engine and measured overhead

The engine is hybrid Python + Rust via PyO3. Latency-sensitive paths (anti-injection regex, PII redaction, forensic hashing) are in Rust. Measured median overhead on the full four-domain pipeline is 6.25 µs (P95 7.04 µs, P99 7.29 µs). The benchmark suite lives in scripts/benchmark.py, with a reproducible Docker environment in docker-compose.benchmark.yml.

SDK

Four primitives, each with async and sync interfaces:

  • GovernedModel — wraps any LLM with automatic PII redaction, audit trail and event emission; Ollama and OpenAI adapters built in
  • GovernedData — enforces residency zones, classification and PII scrubbing on every ingest and query
  • GovernedAgent — wraps MCP / tool-calling agents with firewall, loop detection and PII layers in-process
  • ComplianceKit — EU AI Act risk classification, gap analysis and structured report generation

Top-level imports: from admina import GovernedModel, GovernedData, GovernedAgent, ComplianceKit. py.typed marker and full type hints.

Proxy

FastAPI proxy on port 8080 with JSON-RPC 2.0 passthrough on POST /mcp and bidirectional inspection. REST API: POST /api/v1/validate, POST /api/v1/audit, POST /api/compliance/classify, GET /health, GET /governance/status, GET /api/stats, GET /api/forensic/verify. Rate limiting via Redis, session tracking, CORS middleware. Internal service ports (ClickHouse, Redis) are bound to 127.0.0.1.

Plugin system

Nine interfaces with entry-point discovery and manual registration: BaseModelAdapter, BaseDataConnector, BaseGovernanceGuard, BaseComplianceTemplate, BaseTransportAdapter, BaseForensicStore, BaseAuthProvider, BasePIIEngine, BaseAlertChannel. Reference implementations shipped for Ollama, OpenAI, ChromaDB, filesystem, MCP, HTTP/REST, MinIO, API key, spaCy + regex, log, webhook, EU AI Act template. An optional GuardrailsAI plugin is available for toxic language, jailbreak, bias and PII detection with local-only inference.

CLI

  • admina init <project> — scaffolds admina.yaml, docker-compose.yml and example agent code
  • admina dev — starts the full governance stack via Docker Compose with live health monitoring
  • admina plugin list | install | create — plugin lifecycle management
  • admina doctor — environment diagnostic (Python, Rust engine, plugin discovery, config validity, infrastructure reachability)

Dashboard

Real-time React single-page app on port 3000. Features: Admina Score (0–100 live runtime composite), live event feed via WebSocket, EU AI Act compliance gap view, data sovereignty map, model status. Dedicated OISG (Open / Intelligent / Secure / Governed) adequacy widget: 2×2 quadrant map, total score and 20-criteria checklist, placed in an “Instance Configuration” section that distinguishes static capability assessment from live runtime metrics. Endpoints: /api/dashboard/score, /feed, /compliance, /sovereignty, /infra, /models, /oisg. nginx reverse proxy routes /api/* to the proxy container. Commit hash injected at build time for version traceability.

Integrations

  • LangChain — AdminaCallbackHandler governs every LLM call and tool invocation in-process
  • CrewAI — admina_step_callback and admina_task_callback for multi-agent governance
  • n8n — community nodes package n8n-nodes-admina with AdminaGovern, AdminaAudit, AdminaDashboard
  • OpenClaw — admina-governance skill routing agent actions through the proxy before execution
  • Cheshire Cat AI — three Python hooks (agent_fast_reply, before_cat_sends_message, before_cat_recalls_memories) with a sidecar setup script

Tooling and quality

Environment managed with uv (uv sync --group dev) and uv.lock for deterministic dependency resolution. Python 3.11+ (pinned via .python-version). ruff for lint and format, bandit for security linting, safety for dependency vulnerability scanning. pytest with a coverage gate at 70%. GitHub Actions for CI (lint, test, coverage, security scan) and a release workflow with multi-platform wheel builds (Linux x86_64 / aarch64, macOS x86_64 / arm64) and automated package publishing on tag.

Install

For the full stack with dashboard and infrastructure via Docker:

git clone https://github.com/admina-org/admina.git
cd admina
./scripts/bootstrap-secrets.sh
docker compose up --build

Dashboard at http://localhost:3000, API docs at http://localhost:8080/docs.