Linux
LiveOne script bootstraps Docker or Podman, waits for health, then hands you the lifecycle commands.
scripts/install.sh
Install targets
Run it where you want — within reason
Nebula-COS is being prepared for one-script installs. Linux and Windows (via WSL2 / Docker Desktop) are the v1 targets; macOS is explicitly not. The installer bootstraps Docker or Podman, waits for health, and gives you lifecycle commands once release-gate evidence is complete.
Windows
LiveRuns via WSL2 / Docker Desktop. Same lifecycle, same health-gated install.
scripts/install.ps1Lifecycle commands: install, start, upgrade, doctor, logs, stop, uninstall.
macOS is explicitly not a v1 target.
Read this
Local-first ≠ local LLM
Self-host the backend
Local models · off by default
Local-first means the backend and your data live on your machine — bound to 127.0.0.1, stored in SQLite you own, backed up redacted, with outbound requests restricted by an egress allowlist.
It does not mean the model runs on your hardware by default. Out of the box, chat calls a hosted OpenAI-compatible provider you choose (DeepSeek v4 is suggested in the setup wizard; OpenAI and any compatible endpoint work). Running models locally is an opt-in lane — a managed llama.cpp container, off by default, with no model bundled, and chat advertised disabled in the alpha.
The privacy story COS honestly tells is: your backend, your data, your allowlist — not on-device inference by default.
Hardware planner
Hardware tiers
The read-only hardware planner detects a tier and recommends a profile. It recommends — it doesn't auto-provision.
mini-pc-cpu-only
The floor — runs the backend + hosted-provider chat fine.
Default install runs here
mini-pc-modest
A little more headroom for the local-model and roadmap lanes.
Heavier lanes optional
desktop-gpu
A GPU opens up the opt-in local inference lane.
Heavier lanes optional
workstation-pro
The most capable detected tier for heavier local lanes.
Heavier lanes optional
apple-silicon
Detected as its own tier for tier-aware recommendations.
Heavier lanes optional
Five tiers from services/hardware.py. The planner is informational — it recommends, it doesn't auto-provision. Heavier lanes (local models, and the roadmap voice/memory stacks) are where tier matters; the default hosted-chat install is light.
Model lanes
Honest framing of the model picture
There is no hardcoded provider in the backend — you supply it at setup. The DeepSeek default is a pre-filled wizard value, not a baked-in dependency.
01
Live
Default chat brain
A hosted OpenAI-compatible provider (DeepSeek v4 suggested in the wizard, user-supplied). The egress allowlist defaults to api.deepseek.com + api.openai.com.
user-supplied provider · egress allowlist
02
Optional · off by default
Local model lane
A managed llama.cpp container — opt-in and off by default, with no model bundled.
managed llama.cpp container · no model bundled
03
Roadmap
Voice models
Parakeet STT, Kokoro TTS over LiveKit — the target voice stack, scaffolded but idle in the alpha; not a running lane.
Parakeet STT · Kokoro TTS · LiveKitThere is no hardcoded provider in the backend — you supply it at setup; the DeepSeek default is a pre-filled wizard value, not a baked-in dependency.
Deployment shapes
Install paths at a glance
Three honest deployment shapes. No "full local out of the box" claim — anywhere.
Hosted-provider
LiveDefault. Backend on your box, chat via your chosen API over the allowlist; lightest, works on the mini-PC tier.
Local-models
Optional · off by defaultOpt-in. Add the llama.cpp overlay to run inference on your own hardware; off by default, you bring the model.
Voice / pairing / artifacts overlays
RoadmapAdditional opt-in surfaces, each gated.
Get started
Join early access, then own it.
Nebula-COS is being prepared for one-script installs on Linux and Windows (WSL2 / Docker Desktop). macOS is not a v1 target. The public source release and clean install proof are still release-gated.