Skip to content

KruxOS v0.0.3 — third release

Released: 2026-07-15 Status: Early beta — not for production use.

v0.0.3 makes KruxOS a self-contained AI appliance you can run entirely on your own hardware. The headline is on-appliance inference: pull open models from Hugging Face and run them on the box's built-in engine, and on a GPU appliance install the NVIDIA driver and CUDA stack in one click — no shell, no manual .run files. Two more model providers land — Grok (SuperGrok subscription) and OpenAI Codex (ChatGPT sign-in) — alongside a turnkey built-in remote access so you can reach your dashboard from anywhere over your own private network, offline (air-gapped) updates, and a two-tier license model (Personal is free).

This release builds on v0.0.2 — see the v0.0.2 release notes for the prior inventory. Every feature was exercised on the release-candidate appliance image in the pre-tag hardware walk; the couple of surfaces that ship but were not separately runtime-verified this cycle are noted inline. The complete, itemised list of changes is in the project CHANGELOG.

Highlights

  • On-appliance local inference. Search and pull any GGUF model straight from Hugging Face, then run it on the appliance's built-in engine — no cloud round-trip. A unified "Local model" provider covers the on-box engine, Ollama, and any OpenAI-compatible endpoint, with a Recommended badge on the suggested model and per-model management from Settings › Local Models. A one-click VirtualBox import opts the guest into the host CPU's AVX2 for a faster code path.
  • One-click GPU acceleration. On a GPU appliance, Settings › Hardware › Install Now fetches the correct NVIDIA driver and a signed CUDA accelerator kit, verifies them, and installs into an isolated A/B slot with a real license attestation shown before anything runs. Once installed, GPU acceleration loads automatically at every boot with fail-closed re-verification (any mismatch falls back to CPU rather than loading an untrusted module). A new Hardware / Drivers page surfaces detection, install state, and a GPU health row on the health endpoint.
  • Two new model providers. Grok joins as a provider — sign in with your X account and use your SuperGrok subscription (Grok 4.5 and the Composer/Build variants). OpenAI Codex signs in with ChatGPT (OAuth device-code). Both sit alongside the existing Anthropic, OpenAI, and OpenRouter providers.
  • Built-in remote access (Tailscale). Reach your dashboard — and, if you opt in, SSH — from your own devices anywhere. Turn it on from an optional first-boot step or Settings › System: the appliance joins your own Tailscale tailnet (owner devices only, bring-your-own free account) and publishes the dashboard at a valid-HTTPS address. Nothing is exposed to the public internet, and it is off by default.
  • Offline (air-gapped) updates. Appliances without internet access can now be updated from a file: download the release image and its signature on another device, upload both from Settings › Updates, and apply. It goes through the same signed, verified A/B path as an online update — an unsigned or tampered image is rejected and the disk is never touched.
  • Two-tier licensing. A new entitlement model — Personal (free) and Enterprise — with opt-out anonymous activation telemetry and a read-only license-server admin portal for registrations and usage.
  • Secrets agents can use but never see. Store a scoped, host-bound third-party API key from the dashboard or CLI; agents reference it by name through the new gateway-mediated network.credentialed_request and never see the value.

Added

  • Local models & inference: Hugging Face model search + pull; unified "Local model" provider (on-box engine / Ollama / OpenAI-compatible); Model-license ADVISORY badges on every local model row; community inference-model catalog merged from a contributor-maintained list (hash-verified, additive-only).
  • GPU: one-click "Install Now" driver + CUDA-kit install; appliance-side kit auto-fetch; automatic GPU load at boot with fail-closed re-verification; Hardware / Drivers page; GPU health row on the health endpoint.
  • Model providers: Grok (SuperGrok subscription, sign in with X); OpenAI Codex (ChatGPT OAuth device-code).
  • Remote access & SSH: built-in Tailscale remote access with a guided first-boot step and a Settings card; opt-in SSH with a wizard step that walks you through creating an SSH key.
  • Updates: offline (air-gapped) update from a file, via the dashboard or kruxos update apply <path>.
  • Licensing: Personal/Enterprise tiers; activation telemetry (opt-out); offline-activated appliances register with your account; license-server admin portal.
  • Capabilities: filesystem.edit and richer search / search_content aggregation; network.credentialed_request (use a vault secret without seeing it); add a scoped, host-bound secret from the dashboard or CLI; kruxos vault status.
  • Dashboard & CLI: an Alerts page; a persistent license banner + nav attention marker; Audit/Activity rows now show which CLI initiated each action; restart the gateway from the dashboard; Max Concurrent Code Sessions is now an operator control; configurable pack runtime; kruxos status renders the full health surface (services + resources).

Changed

  • Model provider setup is now one guided "Local model" provider instead of separate entries; the Ollama provider drops the /v1 preset and adds a model field.
  • Same-model context compaction now works for every provider (previously a no-op without Anthropic connected).
  • The first-boot banner and every derived version surface now read the real version from the OS release metadata, so the appliance no longer shows a stale version string.
  • Enterprise upgrade prompts are contact-driven (talk to us) rather than pointing at a self-serve pricing page.

Fixed

  • A resumed session waiting on an approval no longer silently drops the client's next request, and approval holds survive a brief disconnect.
  • Reconnecting a Service Proxy account with a different login now purges the previous account's synced data.
  • OTA image downloads retry instead of failing on a single transient error; the Docker container warns at startup when its sandbox is degraded.
  • Numerous public-docs accuracy fixes across install, monitoring, model providers, GPU drivers, packs, and the SDK guide.

Security

  • Supervision no longer uses a network port or a shared passphrase — it rides a root-only local control socket authenticated by Unix peer credentials.
  • Scheduled jobs now run with the full per-agent scope (workspace / filesystem / process), and an agent can only manage its own scheduled jobs.
  • Per-agent gateway-state isolation across agents and sessions; session buckets are garbage-collected on an idle TTL.
  • A structured "blocked syscall" error is surfaced on a seccomp setup failure rather than an opaque crash.

Upgrading

From a v0.0.2 appliance, update in place from Settings → Updates (online), or upload the v0.0.3 image + signature for an offline update. /data — your agents, vault, licenses, models, and settings — is preserved across the A/B update. For a fresh install, download the whole-disk image from GitHub Releases and write it to your target disk.