Agentic OS
Local. Private. Free.

NovaClaw presents you a party of persistent AI agents. Give them quests - they work and report back. Build software, automate tasks, run marketing campaigns and chat with clients, while running unattended for days. Robust, unkillable and self-sufficient - the main features of Nova, which can deploy itself over a fleet of machines, powered by small cheap reliable local models.

NovaClaw's home screen: Nova's status card with running agents, tokens per second and memory, above a tile launcher for Models, Recipes, Files, Terminal, Debug, Swarm, Help and Settings, and a bar that asks anything.

β—† What makes it NovaClaw

πŸ‘‘ A party, not a lone assistant

Nova β€” the CEO agent β€” hires officers onto real project folders: persistent, named agents, each responsible for its own part of the project, talking to each other like colleagues. The workflows they set up hold for the long term β€” not disposable workers that fire and forget. Set an agent a goal and it runs unattended, exiting only when its work passes review.

🧡 An OS, not a chatbot

Agent sessions are processes: they hand over results a parent can join, and show up in a task manager you can suspend, resume, or kill. Home is an app launcher, not a wall of text.

🏠 Small local models, by design

Point it at any OpenAI-compatible endpoint β€” vLLM, llama.cpp, LM Studio, Ollama. The harness keeps context compacted, recovers broken tool calls, and reconnects to its model with patient backoff β€” so cheap small models hold up for days of unattended work.

πŸ”’ Private by architecture

The detach from opencode stripped every cloud service β€” share egress, bundled OAuth plugins, the hosted provider. Chats, code and knowledge never egress; airgap mode and an unplugged cable are a supported way to run, not an afterthought.

πŸ›‘ Unkillable

The zip ships with its own watchdog: a dead server restarts itself, the UI reconnects forever, and agents recover into running state even after a reboot. Config is a local database, editable at runtime β€” a working agent can repair its own instance.

🌐 A fleet, not a server farm

Instances discover each other on your LAN or a public DHT β€” we run no relay, no directory, no server. Agents call agents across machines, and every peer carries a signed honesty ledger: what it promised against what it delivered.

β—† Why NovaClaw?

OpenClaw, the popular open-source assistant, is one personal AI you message from your chat apps β€” NovaClaw asks a different question: what if your agents were a party that holds together without you?

NovaClawOpenClaw
What it is A party of persistent named agents organized like a company β€” Nova hires officers onto projects they keep: long-term workflows, not temporary irresponsible workers, each with its own chats and memory A personal assistant you message from your chat apps; extra agents are configured peers
Getting started Unpack one portable zip, double-click. No Node, no npm, no API keys Node.js 22+, an installer and onboarding wizard, an API key; on Windows the most Linux-compatible gateway runtime is WSL2
Interface Desktop app + web UI β€” the party roster, chat transcripts, files, terminal; from a laptop or a phone browser Messaging-first β€” WhatsApp, Telegram, Slack, Discord β€” plus a TUI and web chat
Models Small, cheap local models are the design target β€” the harness compacts, recovers and reconnects so they hold up Cloud frontier APIs are the onboarding path; local providers are options too
Your conversations Never leave your machine; unplug the cable and keep working Ride third-party messaging clouds by design
Crash reports On by default, and said out loud β€” scrubbed crash diagnostics only, never your conversations or files. One switch turns it off, and airgap mode disables it outright. Self-hosted, so there is little to disclose; we spell ours out regardless
Between machines Instances find each other on the LAN or a public DHT; agents call agents across machines; we run no server One gateway per host, reached over tunnels (e.g. Tailscale)
Self-healing Ships with its own watchdog; a dead server restarts and agents recover into running state β€” even after a reboot The gateway stays alive as an OS service you set up (systemd/NSSM)