reefyai/reefy

The platform for AI computers

What it solves

Reefy OS is designed to simplify the deployment and management of AI infrastructure on hardware you own. It removes the complexity of "driver wrangling" and manual machine configuration, allowing users to turn PCs, workstations, and servers into a unified fleet of AI computers that are easy to deploy, manage, and recover.

How it works

Reefy uses a minimal, immutable host OS based on Buildroot and a desired-state architecture. It separates the host OS from the application layer, where AI runtimes, frameworks, and models are delivered as versioned container images.

Key technical components include:

  • Accelerator Support Bundles: Build-matched drivers and firmware for NVIDIA, AMD, and Intel (GPU/NPU) that are synced with the kernel version to prevent mismatches.
  • A/B Firmware Updates: A health-gated system with two EFI slots and a hardware watchdog for automatic rollback if an update fails.
  • Encrypted Storage: A stack using LUKS2, LVM thin pools, and XFS to provide encrypted, snapshot-ready application data.
  • Centralized Dashboard: A single interface for fleet-wide monitoring (CPU, GPU, memory, network) and application control via an authenticated MQTT connection.

Who it’s for

It is intended for developers and operators who want to run AI workloads on their own hardware (home labs, edge systems, or datacenters) without the cloud, while maintaining professional-grade infrastructure management tools.

Highlights

  • Multi-Hardware Support: Unified platform for Intel, AMD, and NVIDIA accelerators.
  • Cloud-Independent: Applications start from local cached state and do not require a cloud connection to boot.
  • ** uma Immutable Host**: Read-only SquashFS base with no general-purpose package manager to reduce attack surface.
  • Automated Recovery: Hardware-watchdog and A/B slot recovery for unattended machines.
  • Integrated Monitoring: Out-of-the-box time series metrics for system and accelerator performance.

Related

  • Project
    GCWing/OpenBitFun

    An open-source desktop workspace for AI agents featuring a Rust runtime and multi-device synchronization to manage complex, long-running AI workflows.

  • Project
    DashAISoftware/dashAI

    A graphical toolbox for training, evaluating, and deploying state-of-the-art AI models with support for various hardware accelerators.

  • Project
    lexmin0412/dify-app-hub

    A management platform and optimized user interface for Dify applications, providing a professional front-end for Chatflow and Workflow AI agents.

  • Project
    cloudflare/cloudflare-os

    An AI productivity environment that lets users create and customize private, sandboxed applications called Gadgets using AI agents, featuring a capability-based security framework.

  • Project
    skypilot-org/skypilot

    SkyPilot is an open‑source system that lets AI teams launch, schedule, and scale training, inference, and other AI jobs on any combination of Kubernetes, Slurm, or public‑cloud GPU/TPU resources. You describe a job in a short YAML/Python spec (resources, code, setup, run); SkyPilot automatically picks the cheapest available provider, provisions the machines, syncs your code, runs the job, and shuts down idle resources. Features include auto‑failover, bin‑packing, gang scheduling, multi‑cluster control, production‑grade endpoints, and sandboxed execution of untrusted LLM‑generated code. Install via `uv pip install "skypilot[aws,gcp,azure,kubernetes,…]"` and launch with `sky launch <spec>`. It supports dozens of clouds and integrates with major frameworks (PyTorch, DeepSpeed, Jax/TPU, vLLM, Airflow, Jupyter, etc.), making AI compute portable and cost‑effective.