Private AI. In a Box You Control.
The SAFi Appliance is an AI system engineered to run local AI models governed by an organizational policy. It comes pre-loaded with three open-source reasoning models (Qwen.3.8 27B, Qwen.3.6 35B-3B MoE, and Gemma4 26B-A4B) and three ready-to-use IT agents: Sysadmin, Security Engineer, and Software Engineer.
All vendor-baked censorship and arbitrary safety refusals have been stripped from the models. Runtime safety is strictly governed by your organization's policies instead, giving you total sovereignty over your AI with zero cloud egress.
- 64GB Unified Memory Debian Appliance (Zero Cloud Egress)
- 3 Pinned Models: Qwen.3.8 (27B), Qwen.3.6 (35B-3B MoE) & Gemma4 (26B-A4B)
- Turnkey Sysadmin, Security & Dev Agents
- Policy-Governed Alignment (Zero Vendor Censorship & False Refusals)
- Cryptographic Hash-Chained Audit Ledger
Deployment & Operations
How the appliance works
Get up and running in less than 15 minutes. The appliance is shipped ready for you to connect to your network and log in, with all models and agents pre-configured to work out of the box.
Network Setup & First Login
Connect power and Ethernet: a simple console prompt configures your network (DHCP or static IP) and immediately displays your dedicated local web dashboard and CLI endpoint.
Client or Local Access
Connect remotely from any workstation on your LAN (Windows, macOS, or Linux) via a web browser or SSH. Or, connect a monitor and keyboard directly to the unit to use it as a dedicated station.
Client-Side Sandboxing
When using the terminal harness (safi-cli), agents execute tools locally on the client machine inside unprivileged Linux Bubblewrap (bwrap) sandboxes. Commands are strictly restricted to the project workspace with zero access to your underlying host files, SSH keys, or cloud credentials.
Audience & Philosophy
Who the appliance is for
The SAFi Appliance is built for IT directors, sysadmins, security analysts, and developers who need AI leverage but cannot, or choose not to, send operational data to public cloud models due to regulatory requirements, compliance mandates, or internal privacy standards.
Companions, Not Replacements
SAFi agents are designed as assistants and force multipliers, not autonomous replacements for your staff. The system assumes the user understands the underlying infrastructure and can review, guide, and verify the work being performed.
Strict Privacy & Compliance
Ideal for defense contractors (CMMC / ITAR), healthcare (HIPAA), finance (SEC / FINRA), and enterprise teams whose security policies forbid exposing internal configurations, private IPs, or source code to public LLM vendors.
The IT Agent Triad
Familiar terminal workflow. Server-enforced governance.
If you are familiar with terminal coding harnesses such as Claude Code or OpenCode, you will feel right at home with SAFi CLI. The key difference: while traditional harnesses lack centralized control, SAFi cleanly separates governance from execution. The appliance server hosts your private AI models and enforces policy guardrails, while your local workstation (the client) executes all agent tools safely inside unprivileged Linux Bubblewrap sandboxes.
Client-Side Sandboxing & Workspace Isolation
All tool execution runs on your local client: When you run safi-cli on your machine, tool commands (Bash, file edits, git) execute locally on your workstation, isolated inside a Linux Bubblewrap (bwrap) sandbox. Whichever directory you launch the agent from becomes its locked workspace: agents cannot access or modify anything outside that path.
Controlled cross-session memory: The only exception is the Sysadmin Agent, which is exclusively granted read access to a local inventory.json file in your client's configuration directory. Define your infrastructure map once on your client, and the Sysadmin retains that operational knowledge across sessions without exposing it to other agents.
The Sysadmin Agent
Hands-On Sysadmin Automation
- Operates Windows, Linux, and cloud infrastructure using native Bash, PowerShell, SSH, and Ansible tooling under strict policy guardrails.
- Production-ready for enterprise Linux distros, Windows Server, Active Directory, and Microsoft 365.
- Hardened Remote Access: SSH key authentication is strongly recommended to connect to remote servers (with helper scripts included to simplify setup), while standard credential logins remain available when needed.
- Automates service diagnostics, log triage (journalctl, Windows Event Viewer), and fleet maintenance routines.
- Tackles tedious operational ticket backlogs without requiring public cloud AI subscriptions.
The Security Engineer
Vulnerability Triage & Hardening
- Private CVE triage: Ingests Nessus and Qualys scan exports locally without uploading internal hostnames, IPs, or vulnerability maps to the cloud.
- Web & OS server hardening: Audits Nginx, Apache, SSH, and systemd units directly against CIS Benchmarks.
- Firewall & network audits: Reviews iptables, UFW, NAT tables, and cloud security group configs for accidental exposure.
- Credential hunting: Catches hardcoded API keys, exposed
.envfiles, and improper file permissions across repositories. - Zero false refusals: Runs authorized security audits under your organization's policy without being blocked by third-party AI safety filters.
The Software Engineer
Repo Engineering & Glue Code
- Inspects, refactors, and modernizes internal scripts, webhooks, and automation glue code across tech stacks.
- Generates comprehensive unit, integration, and regression test harnesses.
- Performs automated code reviews and builds technical documentation from source.
- Confined strictly to the active repository workspace with zero access to private host credentials.
Build Custom Agents for Your Stack
The IT Triad gets your team working on day one, and you are never limited to just three roles. Define and deploy as many custom, role-specific agents as your workflow needs, each governed by its own SAFi policy charter, tool allow-list, and isolated Bubblewrap workspace.
Local Foundation Weights
The AI models
Pre-loaded, 4-bit quantized, and benchmarked for local silicon. Instruction-following precision with zero vendor refusal heuristics.
Qwen.3.8 (27B)
Dense reasoning workhorse in 4-bit quantization. Tuned for complex multi-step Bash planning, software refactoring, and deep code comprehension.
Qwen.3.6 (35B-3B MoE)
The default engine for SAFi runtime policy audits and Conscience evaluations. Delivers blazing-fast compliance verification, rapid log triage, and instant operational commands.
Gemma4 (26B-A4B)
Specialized in strict instruction following, deterministic JSON schema generation, and structured technical outputs.
Abliterated for Policy Alignment
Vendor-baked refusal guardrails frequently block legitimate IT work like penetration audits, firewall updates, and CVE triage. All included models are abliterated so internal refusal vectors never conflict with your organization's written policy.
Continuous Upgrades & Open Weights
We continuously evaluate emerging open-weight releases and provide updated model packages. You are never locked in: load any GGUF or native local model into the appliance whenever you choose.
Silicon & Performance
The Hardware
Our goal is to keep the appliance price accessible while delivering the high throughput required for interactive agent workflows.
The Unified Silicon Advantage
Traditional GPU servers cost $25,000 to $80,000, draw over 1,000 Watts, and require datacenter air conditioning. The SAFi Appliance uses high-bandwidth unified silicon, delivering fast local inference in a whisper-quiet Mini-PC that draws under 120 Watts from any standard wall outlet.
Guaranteed Hardware Floor
Vendor-agnostic capability contract
- Unified Memory: 64GB guaranteed minimum high-bandwidth unified memory. Shared dynamically across compute and inference cores with zero VRAM offload bottlenecks.
- Compute & Acceleration: 16-Core / 32-Thread high-efficiency processor with dedicated hardware acceleration for local tensor and matrix math.
- Storage: 1TB high-speed PCIe Gen4 NVMe SSD pre-loaded with Debian OS, verified models, and cryptographic audit database.
- Enterprise Networking: 2.5 GbE Ethernet port + Wi-Fi 7. Plugs directly into your isolated management VLAN.
- I/O & Display: Dual high-speed USB4 / Type-C ports, HDMI, and DisplayPort for direct local console access.
- Power & Acoustics: Sub-120W peak power consumption. Compact aluminum Mini-PC engineered for whisper-quiet desktop or server shelf placement.
- Hardware Warranty: 1-Year hardware replacement warranty included with advance dispatch.
Pre-Installed Software Stack
Production-ready from first boot
- Hardened SAFi OS: Debian-based appliance system with kernel tuning, hardware acceleration, and Bubblewrap sandboxing.
- 3 Pre-Loaded AI Models: Qwen.3.8 (27B), Qwen.3.6 (35B-3B MoE), and Gemma4 (26B-A4B) benchmarked and quantized on NVMe.
- First-Boot Setup: Simple console prompts for network configuration, followed by web browser and CLI endpoints.
- Local HTTPS: Automated SSL/TLS certificates generated locally with internal CA.
- Identity & Auth: Native Active Directory / LDAP and SCIM 2.0 provisioning.
- True Air-Gap: Pull the network cable: all models, agents, and governance tools run 100% locally.
Hardware Guarantee: Specifications listed above represent our guaranteed capability floor. We select and calibrate the silicon to guarantee performance parity, and we reserve the right to ship equivalent or better hardware as components evolve.
Enterprise Control & Audit
The Governance
Governance in SAFi is an immutable, server-enforced audit and policy engine. Every agent action is governed by your rules and signed to a cryptographic ledger.
Every AI Agent in SAFi Is Governed by a Policy
In SAFi, every agent behavior and action is governed by a policy written by the individual or organization that owns the system. In the appliance local AI models, the built-in safety and alignment rules set by the AI maker have been removed so there is no conflict between the policy and model built-in rules. Most cloud-based models will refuse to do penetration testing or cyber attack simulations because of built-in rules. In SAFi, all that is controlled through the policy.
The Appliance includes three editable policies for the sysadmin, software, and security engineer agents that can be edited to fit your needs.
The governance framework that makes SAFi possible is based in thousands of years of philosophical thinking, starting from Aristotle to Thomas Aquinas on how we understood agency for thousands of years. If you want to read more about the framework behind SAFi, visit the main project website or inspect the GitHub repository. SAFi is 100% open source.
Input Threat Defense
Phase Zero Gate
Screens raw user prompts deterministically before any model is invoked. Scans for prompt injections, jailbreaks, and sensitive PII. If a threat is detected, it short-circuits immediately with zero model tokens burned.
Deterministic Outcomes
Closed-Loop Enforcement
SAFi operates as a closed-loop system. The AI model can only propose actions, with zero direct execution authority. A deterministic enforcement layer decides if actions are permitted based on policy allow-lists, parameter constraints, and output audits. Every action is guaranteed to produce an audit record or it fails closed.
Cryptographic Audit Trail
Immutable WORM Ledger
Every prompt, draft, tool proposal, evaluation score, and execution verdict is cryptographically signed into an append-only, tamper-evident hash chain accessible via the local appliance dashboard.
Enterprise Oversight & Regulatory Compliance
Designed for regulated enterprises: SEC 17a-4 & FINRA 3110 (tamper-evident WORM audit records), HIPAA & CMMC (zero cloud egress behind local firewalls), and EU AI Act & ISO 42001 (continuous audit trails and human supervisory oversight).