# rzfz.ai

> rzfz.ai builds local AI infrastructure: the rzfz.ai Stack — the integrated
> open-source stack for local AI (chat, workflow automation, knowledge-graph
> RAG, personal and coding agents, document OCR, PII redaction) — and turnkey
> bundles (Box + Stack, Cloud + Stack, own hardware).

- Current release: 2026.09-ga (2026-09-21), CalVer, monthly cycle
- rzfz.ai Subscription: 799 EUR net per server/box/VM per year
- Supported OS: Ubuntu 26.04 LTS. Production inference: llama.cpp via Vulkan (GGUF models)
- Source: https://github.com/rzfz-ai/rzfz-ai-service-stack
- Services: AI-driven software engineering and integrations by razzfazz.ai GmbH; consulting and digitalization delivered by SEQIS.ai (SEQIS Group).

## Pages

- [Stack](https://rzfz.ai/stack/): what the rzfz.ai Stack is
- [Modules](https://rzfz.ai/stack/#architektur): all core + optional modules, live in the Stack Map
- [Features](https://rzfz.ai/stack/#features): the seven local-AI USPs
- [Hardening](https://rzfz.ai/stack/#haertung): SSO/RBAC/MFA, NIS2 + ISO 27001, SBOM per release
- [Enterprise](https://rzfz.ai/enterprise/#subscription): the three tiers — community edition, rzfz.ai Subscription, optional services
- [System requirements](https://rzfz.ai/stack/#systemvoraussetzungen): deployment profiles, honest limitations
- [Releases](https://rzfz.ai/stack/#releases): release history
- [Bundles](https://rzfz.ai/bundles/): Box + Stack, Cloud + Stack, own hardware
- [Engineering & consulting](https://rzfz.ai/beratung/): AI-driven software engineering by razzfazz.ai GmbH; consulting and digitalization via sister company SEQIS.ai

## Modules

### Chat

- Open WebUI (stable) — Local chat interface for every model in the stack

### Workflow automation

- Dify (stable) — Workflow and agent platform — visual orchestration for LLM applications
- Crawl4AI (experimental) — Web crawler for LLMs — extracts web content as clean markdown for RAG pipelines

### RAG & company brain

- SearXNG (stable) — Meta search engine — private web search without tracking, also usable as an agent tool
- Cognee (experimental) — GraphRAG company brain — one central memory for agents, chat and workflows
- LightRAG (experimental) — Lightweight GraphRAG framework — an alternative to Cognee for smaller knowledge bases
- Onyx (experimental) — Enterprise search and RAG chat over connected internal data sources

### Agents

- Agent Manager (experimental) — Agent portal — launch, open and watch personal AI agents in one pane
- MCP Manager (experimental) — MCP manager — registration and permissioning of MCP servers for agents
- Hermes (experimental) — Personal Python agent with memory and terminal UI — provisioned per user
- Moltis (experimental) — Personal agent server in Rust with a Matrix gateway — provisioned per user
- OpenHands (experimental) — Autonomous coding agent — plans, writes and tests code inside a sandbox
- Paperclip (experimental) — Agent orchestrator — organizes teams of agents with tasks, budgets and an audit log
- Codex (experimental) — Codex — OpenAI's terminal coding agent, bundled in the coding agents workspace
- OpenCode (experimental) — OpenCode — an open-source AI coding agent for the terminal, bundled in the coding agents workspace
- gsd-pi (experimental) — gsd-pi — a spec-driven coding agent, available as a user-defined install in the coding agents workspace
- User Defined Coding Agents (experimental) — User-defined coding agents — free choice of your own coding-agent tool in a persistent workspace

### Documents

- Gotenberg (stable) — Document conversion — HTML, Office and Markdown to PDF via a REST API
- Docling (experimental) — Document parser — converts PDFs and Office files into structured text for RAG
- Presidio (experimental) — PII detection and redaction in text and documents before further processing
- Apache Tika (experimental) — Content extraction — text and metadata from over a thousand file formats
- Stirling PDF (experimental) — PDF toolbox — merge, split, convert and edit PDFs
- Paperless-ngx (experimental) — Document archive — scans, indexes and makes paper documents and PDFs searchable

### Security & identity

- Authentik (core) — Identity provider — SSO, RBAC and MFA in front of every module
- Docker Socket Proxy (core) — Docker socket proxy — restricted, read-only access to the Docker API for modules that need it
- Wazuh (experimental) — SIEM/XDR — intrusion detection, file integrity and compliance evidence for the box
- OpenUEM (experimental) — Endpoint management — inventory, software rollout and remote assistance for end devices
- Vaultwarden (experimental) — Password vault — self-hosted, Bitwarden-compatible credential management
- Infisical (experimental) — Secrets management — central handling of API keys and environment variables for the stack

### Operations & monitoring

- Backup Service (core) — Backup service — automated, encrypted backups of every database and volume
- Settings Portal (core) — Settings portal — central interface for configuring the rzfz.ai stack
- Help Center (core) — Help center — locally hosted documentation and support articles for the stack
- Start Portal (core) — Start portal — landing page linking to every unlocked module in the stack
- Licenses (core) — License management — activation and status of unlocked modules and bundles
- Autoheal (core) — Autoheal — watches container health checks and restarts unhealthy containers automatically
- Komodo (stable) — Deployment and monitoring tool for the system's Docker Compose stacks
- Observability (experimental) — Observability — metrics, logs and alerts for every container in the stack

### Infrastructure

- Caddy (core) — Reverse proxy and TLS — the single external entry point of the stack
- PostgreSQL (core) — Relational database — persistent storage backing modules like Open WebUI, Dify and Authentik
- Valkey (core) — In-memory data structure store — caching and message queues for the stack
- SMTP Relay (core) — SMTP relay — outbound mail delivery for system notifications
- LLM Manager (stable) — LLM manager — one console for every language model on the box: deploy, watch, account for usage
- LLM Worker-Agent (stable) — Worker agent — starts and supervises the inference engines on every machine in the fleet
- LLM Model Registry (Zot) (stable) — Model registry — the box's built-in OCI registry for model weights and engine images
- GPUStack 0.7.1 (optional legacy backend) (stable) — GPUStack 0.7.1 — the optional legacy backend, registered behind the LLM manager
- Speaches (stable) — Speech input and output (STT/TTS) — local transcription and speech synthesis behind an OpenAI-compatible API
- Gitea (stable) — Self-hosted Git — repositories and CI hooks for internal projects
- Matrix (experimental) — Matrix homeserver — encrypted chat and bot protocol for agent notifications

## Use cases in production

- AI-assisted annual financial statement review & documentation (Tax advisory, in progress) — A tax advisory firm automates the validation of personnel provisions, revenue and inventory values through intelligent data matching across PDFs, lists and account statements. The notes to the financial statements are prepared in structured form as well — for maximum efficiency and a well-founded basis of argument in upcoming tax audits.
- Support email classification for IT operations (IT services, in progress) — An IT service provider serving many SMBs receives hundreds of log and support emails daily. A local workflow classifies them, separates noise from real incidents and prioritizes what a human needs to see.
- Central RAG knowledge base for production control (Manufacturing, in progress) — A production-control company consolidates internal knowledge — manuals, incident history, process docs — into one central, local knowledge base that staff query via chat.
- 24×7 dev agent raising test coverage (Software development, in progress) — A local coding agent works around the clock on the in-house repository: writing unit tests, steadily raising coverage and opening merge requests — with no code ever leaving the infrastructure.
- Local image analysis in patient care (Healthcare, in progress) — Patient data processing with local vision models: disease symptoms are detected on photos directly on site — sensitive health data stays entirely within the facility.
- Framework for digitalization consulting (Consulting, in progress) — A structured consulting framework guides digitalization projects: process discovery, assessment and prioritization are supported by local AI workflows.
- Local coding with on-box models (Software development, in production) — Developers use coding models directly on their own hardware — completion, reviews and refactorings without external API keys and without code leaving the premises.
- Risk assessment of requirements (Software development / QA, in progress) — Requirements are assessed locally for business risk, failure probability and complexity — informing how intensively each requirement gets tested.
- Finding TARIC goods codes via RAG (Foreign trade / customs, in progress) — The TARIC database as a local RAG: matching import/export goods codes are found from a plain description of the goods — replacing tedious manual search through the nomenclature.
- Document extraction and redaction of sensitive data (Financial sector, in progress) — Inbound customer documents are extracted locally and sensitive data is redacted automatically before further processing — built on the rzfz.ai Stack, awarded the Constantinus Award 2026.

## Deployment profiles

### rzfz.ai Box

AMD Strix Halo (Ryzen AI MAX+), 128 GB unified memory

Requirements:
- Ubuntu 26.04 LTS (pre-installed)
- Network access, ports 22/80/443

Limitations:
- AMD-only inference — no NVIDIA/CUDA path on the box
- vLLM currently unsupported on Strix Halo (production: llama.cpp via Vulkan)
- The 32 GB container budget limits how many modules run concurrently
- One intensive agent loop at a time — sizing rule ~1 box per 6–10 employees

### Own server / VM without GPU

Any Ubuntu 26.04 host, ~16–24 GB RAM

Requirements:
- Ubuntu 26.04 LTS, kernel 7.x
- Docker Engine + Compose v2

Limitations:
- CPU inference is slow — fine for light or batch workloads, not for interactive agents
- As a control plane it delegates inference to GPU workers (boxes or cloud)

### NVIDIA/CUDA server

Ubuntu 26.04 host with NVIDIA GPUs — own hardware or GPU cloud

Requirements:
- Ubuntu 26.04 LTS
- CUDA-capable NVIDIA GPU(s)
- Docker Engine + Compose v2

Limitations:
- Inference via llama.cpp on CUDA, models in GGUF format
- The VRAM ceiling is a budget question, not an architectural one

### Cloud instance

Any Ubuntu 26.04 cloud instance — with GPU as worker, without GPU as control plane

Requirements:
- Ubuntu 26.04 LTS
- Docker Engine + Compose v2

Limitations:
- Data sovereignty depends on the cloud provider — for strictly local scenarios choose the box or own hardware

## Releases

### 2026.09-ga (2026-09-21)

Your models, managed

- LLM Manager and console — deploy language models from one web console: from the curated catalogue or straight through the Hugging Face search, plus a fleet page for workers and nodes, the model hub, API keys with per-key limits and token metering, a playground for chat, embeddings and rerank — and a dashboard whose curves follow the measurements.
- One OpenAI-compatible endpoint — GPUStack is no longer the product but a backend behind it. The address you already use keeps answering — now from the LLM Manager, with its own keys.
- Agent Manager — the fourth app of the one-identity family: a per-user agent portal that shows instances rather than types, opens on the agent's own interface and lets users name their agents; administrators control start, stop and memory for every agent with an audit trail.
- Start portal redesigned — shortcuts extended: favourites and categories on the rail, tiles for every new module, and curated one-click tiles: a chat persona with a preset model and prompt, a Dify app, or an external assistant opened with a prefilled prompt — defined once, scoped to Authentik groups.
- Settings portal redesigned — at its own address, with the information-first layout of the LLM console, a master-detail pane for settings, and toggles for the new module profiles.
- Help Center and licence overview redesigned — two-level navigation, local full-text search and mirrored upstream documentation for every module, plus our own guides for the LLM Manager and the Agent Manager; the licence page shows tier cards, licence families and the full licence texts, which now ship inside the product.
- One look across all apps — the same signature and the same dark navigation rail in the start portal, Settings, Licences, Help Center, LLM console and Agent Manager.
- Wazuh and OpenUEM — both experimental, off by default: SIEM/XDR with file-integrity monitoring, log correlation and compliance mappings; endpoint inventory, software deployment and remote assistance for the fleet.
- Chat and workflows brought up to date — Open WebUI 0.11 with a redesigned interface, human-in-the-loop tool approval and sub-agents; Dify 1.17 with dataset-scoped Knowledge API keys, workflows that can ask a person a question, and skill management. Plus security updates for Authentik, Gitea, Valkey, Crawl4AI, Element Web and Vaultwarden.
- First install hardened (2026.09-ga.1) — a stalled model download resumes instead of starting over, post-install says sooner when a model has finally failed instead of waiting out its full budget, and the readiness check understands a box that is online.
- Before you upgrade — Anything using your box's model API — the OpenAI-compatible endpoint for language models — needs a new key after the upgrade. The address you already use keeps working; only the key changes. Mint one key per consumer in the LLM Manager console and swap it in.

### v2026.08-ga (2026-07-27)

Offline & air-gapped operation: the box runs entirely without internet; installing and updating without internet with the rzfz.ai Subscription.

- Offline & air-gapped — module operation without internet access, and installation and updates without internet with the rzfz.ai Subscription: everything the box needs for it is held locally, with completeness checks before each restart and optional kernel-side egress sealing of the host.
- Apple Silicon inference (optional) — Mac machines can be connected through a gateway as additional inference nodes.
- rzfz.ai Shortcuts — curated mini-apps for business users, managed through the start portal.
- Enterprise documentation portal — access-protected customer documentation at docs.rzfz.ai.
- Domain change in one click — a changed box domain name now propagates fully: SSO provider, login URLs and every service follow automatically.
- Password manager hardened — Vaultwarden stores in PostgreSQL (part of the standard backup), SSO login issues fixed, compatibility with current Bitwarden clients.
- Security and currency sweep — including a critical Gitea hotfix (CVE-2026-59774), updates for Matrix, LightRAG, Valkey and base images; four hardened internal auth paths.
- Continuously hardened across 15 patch releases; public downloads now on GitHub (github.com/rzfz-ai/rzfz-ai-service-stack).

### v2026.07-ga (2026-07-06)

Open-core licensing goes public: Apache-2.0 Community, the rzfz.ai Subscription and coding-agent workspaces

- Open-core licensing, delivered publicly — a free Apache-2.0 Community tier, a source-available tier (BUSL-1.1) under the rzfz.ai Subscription, a live license overview, a public GitHub mirror and a published Community Wiki
- Coding-agent workspaces — personal coding agents are now per-type sandboxed containers (OpenCode, Codex, user-defined) with live web-app preview plus per-user MCP integrations and memory
- Dify 1.15.0 — single sign-on no longer asks you to log in twice

### v2026.06-ga (2026-06-14)

One default model: qwen3.6 at 1M context for chat, coding and vision

- qwen3.6 is the one default model for chat, coding, general tasks and vision at 1M context — no per-role model juggling
- Document Q&A works out of the box — the reranker and chunking defaults that make RAG find the right facts ship pre-configured
- Smoother upgrades — both the standard and the from-2026.04 upgrade paths are validated, with upgrade self-healing and journaling
- Audit-ready compliance tables — structured NIS2 (EU 2022/2555) and ISO/IEC 27001:2022 evidence in the security architecture

### v2026.05-ga (2026-05-08)

Per-user personal agents: Hermes, Moltis and coding agents in the "My Agents" drawer

- rzfz.ai start portal — tile launcher with per-user pinning
- New Crawl4AI module and observability profile (OpenLIT + ClickHouse)
- Open WebUI ↔ Dify manifold pipe

### v2026.04-GA (2026-04-06)

One command from base install to fully configured stack: razzfazz-post-install.sh

- razzfazz-post-install.sh automates GPUStack model deployment plus Open WebUI and Dify setup in one step
- Two experimental RAG modules: LightRAG (graph-aware) and Cognee with FalkorDB
- Encrypted .env snapshots before every configuration change; AMD GPU inference fix

## rzfz.ai Subscription

799 EUR per server/box/VM per year, net

Includes:
- Right to operate the current, patched stack commercially
- At least 4 releases per year (currently monthly) incl. security updates
- Access to the Enterprise documentation (docs.rzfz.ai)
- One subscription per stack installation (server, VM or box)