Nebula · by Sotto Labs · Private alpha

Nebula

Your private AI has a home.

One identity across your computer and phone, designed to remember, help, and act under your control. The public alpha proves the self-hosted trust kernel today; seamless voice, proactive memory, and complete phone continuity are still being qualified.

See what works today → Join early access →
Your computer is home the public core binds locally by default
Your model choice connect a supported provider; local lanes are in qualification
Your permission layer receipts, egress rules, recovery, and explicit gates

Vision keyshot. Current alpha surfaces vary. Voice, local models, and proactive memory are roadmap, gated, or off by default—and labeled precisely throughout.

What it is

What Nebula-COS actually is

Not another chatbot wrapper, and not a finished product. Nebula-COS is the installable, production-grade foundation of the Nebula line: a FastAPI backend with a real trust kernel around an OpenAI-compatible chat core.

A backend you own

Live

Docker Compose, binds to localhost by default, SQLite storage with schema versioning and corruption recovery. The center of gravity is on hardware you control.

A provider you choose

Live

Chat routes to a hosted OpenAI-compatible endpoint you configure at setup; DeepSeek v4 is pre-filled in the wizard, OpenAI and others work too. An egress allowlist restricts where requests can go.

A trust kernel that's real today

Live

Memory ledger, action receipts, external-content firewall, backup/restore, and a capabilities matrix — all live and tested.

Live now

Verified, not aspirational

Every claim below carries a Live tag and an evidence line naming the code behind it. Honesty is the headline.

01 Live

Self-hosted backend

Docker Compose, serving on 127.0.0.1:8080 by default.

under the hood: apps/server/cos + compose.yaml
02 Live

Provider-bridge chat

/v1/turn against any OpenAI-compatible API; DeepSeek v4 default in the wizard.

under the hood: services/provider.py, services/turn.py
03 Live

Memory ledger

CRUD + review states + instruction detection. Honestly: it does not auto-inject into your chat context yet.

under the hood: capabilities turn_context=False
04 Live

Action receipts

Every action run leaves evidence + recovery. Today: the local_note.create action.

under the hood: ActionRun receipts, postconditions
05 Live

External-content firewall

Untrusted content is quarantined, reviewed, and excluded from backups.

under the hood: quarantine + review gate
06 Live

Egress allowlist + sanitizer

Outbound hosts restricted, private IPs rejected, credentials scrubbed.

under the hood: egress allowlist + outbound sanitizer
Roadmap

Honest about what isn't here yet

The headline Nebula pillars — talk-to-it voice, memory that recalls across conversations, and a proactive assistant that acts on its own — are not in this alpha. They live in the private Nebula engine and are being ported. We label them as roadmap everywhere, never as shipped.

Voice loop (LiveKit + Parakeet + Kokoro) — scaffolded, idle by default Agentic memory recall (Journal + RRF retrieval) — not yet built in COS Proactive Pulse / action proposals — not yet built in COS

See the full live-vs-roadmap map →

Abstract diagram of the Nebula-COS layers, from surfaces down to models

Models think. Nebula gives you a place to own what they remember, restrict where they reach, and audit what they do.

Why not just use ChatGPT, Claude, or Gemini? → Install it →
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.

Join early access → Read the alpha path

Managed-install interest is being collected for the alpha; support availability follows release-gate evidence.