The adoption cycle, built on an honest reading: the feature depth already exceeded what an adopter needs, and what was missing was being found, being believed, and working on the hardware people actually own. Systems installed from 0.4.0 onward receive this release through a plain sudo apt update && sudo apt upgrade, and CI keeps proving that survival on every push.
What is new
Believed: the README now answers the question every visitor silently asks - "why not just Ollama?" - front and center, six differences in a table, with the honest closing that localhost-only users need none of it. And the terminal demo stops being acted: the CI pipeline records the real session (asciinema) on every push, asserts the completion is present and the API key absent, and ships the recording; what the README shows is generated, not staged.
Working on real hardware: following the only real inbound signal the project has received (a Strix Halo/ROCm question), AMD is now first-class on the consumer path. First boot completes the ollama ROCm runtime on AMD machines whose GPU was not visible at image bake time - cloud images are baked in GPU-less VMs, so AMD deployments used to fall back to CPU silently. sibilla status, doctor and bench name AMD cards and say loudly whether the ROCm runtime is present. The detection logic is a deterministic CI matrix (datacenter NVIDIA, consumer NVIDIA, AMD discrete, APU, none), and docs/validation/gpu.md is the tester guide with a structured issue form: if you own an NVIDIA card, an AMD card or a Ryzen APU, thirty minutes of your time is exactly the data this project cannot produce alone.
Supply freshness: the ollama pin moves to 0.32.1, the vLLM container is pinned to v0.25.1 (a ":latest" had survived from the PoC), and sha256 digests are now recorded for every entry of the catalog - 161 of them - so model use, model pull, sibilla model import and the air-gapped payload verify any catalog model, not just the CI ones. One finding became policy: recent ollama ships with cloud routing enabled by default and can transparently offload ":cloud" models to ollama.com; every SibillaOS install now pins OLLAMA_NO_CLOUD=1 and CI asserts it, because "nothing leaves the machine" has to include the engine's own cloud.
The config track completes: sibilla apply check audits without changing - drift against the declared profile, a hand-edited Caddyfile (which the next toggle would silently overwrite), the served model against its catalog digest, key file permissions - and exits nonzero on findings, cron-ready.
Local RAG, from feature to recipe: examples/rag-quickstart.py is a complete pipeline in ~100 lines of Python standard library, environment straight from sibilla connect --env, run end to end by CI; LangChain and LlamaIndex configurations ship with the local-endpoint gotchas stated.
The developer kit grows its missing half: sibilla connect --remote generates a deliberately keyless script for the workstation - it prompts for the key on stdin there (never argv, never shell history), refuses to write anything until the gateway accepts the key, then places Continue and aider configs plus an optional project .env. Zed gets an honest print-only section (its keys live in its own credential store).
And the machine now looks like itself: the long-form commands (status, doctor, bench, apply check) wear the emerald oracle style - color only on a TTY, so pipes and CI see byte-identical plain text; the appliance images greet SSH logins with the ASCII oracle eye, the served model and where to start (guest installs stay untouched, asserted); the gateway answers Server: SibillaOS; Open WebUI presents itself as part of the appliance; the wallpaper is redesigned ("quiet corner") and the eye lost its gold glint across every asset - the slit pupil carries the mark alone.
What has been verified
Everything the previous releases established, plus: the recorded demo asserted on every push, the hardware-detection matrix, OLLAMA_NO_CLOUD on both install paths, the config audit catching a tampered Caddyfile and a drifted profile, the RAG example running the documented flow, the remote kit refusing a bad key and writing three files correctly on a good one, the login banner present on images and absent on guests, the Server header, and the hard embedding assert in the pinned VM (the native runners follow the same best-effort policy as chat inference, with engine failures named and our own failures still red). The catalog pipeline itself came out hardened: the updater can no longer truncate its own catalog, signed files are pinned to LF, and lint and build check substance, not just syntax and signature.
Installation
On an existing Ubuntu 24.04 machine:
$ curl -fsSL https://engineering87.github.io/sibillaos/apt/sibillaos-archive-key.asc \
| sudo gpg --dearmor -o /usr/share/keyrings/sibillaos-archive-keyring.gpg
$ printf 'Types: deb\nURIs: https://engineering87.github.io/sibillaos/apt/\nSuites: ./\nSigned-By: /usr/share/keyrings/sibillaos-archive-keyring.gpg\n' \
| sudo tee /etc/apt/sources.list.d/sibillaos.sources
$ sudo apt update && sudo apt install llmd
$ sudo sibilla setupReversible with sudo sibilla remove, CI-verified. From the ISO (amd64): .part files plus SHA256SUMS, reassemble, verify, write to USB. From the cloud image (amd64/arm64): verify against SHA256SUMS-cloud-<arch>, boot with your cloud-init user-data - optionally including an /etc/llmd/profile the machine converges on.
Known limitations
Plain HTTP with mandatory bearer keys by default: enable TLS with sibilla tls before exposing the port. Rate limiting is per-IP; per-key fairness stays deferred with the Caddy-vs-nginx decision recorded. Air-gapped payloads cover the Ollama engine. GPU support on real silicon awaits tester reports (see the call in the README); vLLM on physical datacenter GPUs remains the standing v1.0 debt. The curated catalog only contains permissively licensed, non-gated models.
Licensing
SibillaOS components are licensed under Apache-2.0. The images bundle vLLM (Apache-2.0), Ollama (MIT) and llmfit (MIT); Open WebUI is pulled at the user's request. NVIDIA drivers are installed from the Ubuntu restricted component and are not redistributed. Models are downloaded by the user from Hugging Face under their respective licenses.
Checksums and attestations
See the attached SHA256SUMS files; SBOMs and the installed-package inventory are attached to this release.
What's Changed
- release: v0.7.0 - found, believed, working on real hardware by @engineering87 in #6
Full Changelog: v0.6.0...v0.7.0
