NexusOS is a complete desktop operating system that runs in your browser (or as an Electron app) where the AI isn't a chatbot on the side — it's a kernel-resident service with the same authority as the user.
The AI can write files, build apps, spawn sub-agents, search the web, run code, modify its own skills, and remember your conversations. You watch it work. You can stop it anytime.
It's not a wrapper around ChatGPT. It's an OS built from the ground up for autonomy.
The AI has real authority. Not "suggest a command and you copy-paste it." The AI calls structured tools (write_file, build_app, spawn_agent...) via native function calling and they execute immediately, bounded by a policy engine with hard barriers.
It evolves. The AI writes JavaScript skills, persists them to its own filesystem, and invokes them later. Each skill runs in an isolated Web Worker sandbox. Over time, the OS gets smarter — not because you updated it, but because the AI taught itself new capabilities.
It remembers. Every conversation is indexed into a vector store. Ask "what did we build last week?" and the AI recalls it semantically.
It generates real apps. Not single HTML files. Full app directories: manifest.json, index.html, styles.css, app.js, README.md. Apps appear in the launcher, run in isolated iframes, and can be edited as multi-file projects in the built-in IDE.
| Module | What it does |
|---|---|
78 OS:: actions |
Structured tool surface — files, apps, browser, agents, RAG, git, vision, voice, skills, goals, data analysis, cron scheduling, window sessions |
| 20 AI providers | Z.AI Coding Plan (GLM-5.1/5.2, Turbo, Coding Series), OpenAI, Anthropic, Google, Mistral, NVIDIA NIM, Groq, xAI, DeepSeek, Cerebras, Perplexity, local models (LM Studio, Ollama, Wllama) + failover |
| MCP (Model Context Protocol) | Native MCP client over streamable HTTP — connect any MCP server (filesystem, git, PostgreSQL, browser…) and its tools become native function-calls the DAEMON can drive. Configurable in Settings → MCP Servers |
| SkillForge v2 | AI writes + persists + executes JS skills in a Web Worker sandbox |
| Skill Pack | 29 curated pre-forged skills shipped on boot (data, advanced CSV, dev, web, a mini persistent JSON database, image-info, structured scraping, skill chaining, productivity, AI-assisted) seeded idempotently without clobbering the AI's own customizations |
| AutoPilot | Goal queue with self-prompting — the AI picks goals and works through them autonomously |
| Multi-Agent v2 | Parallel sub-agents with dependency graph + shared workspace + inter-agent messaging |
| AppGenerator | Generates complete multi-file apps (manifest, HTML, CSS, JS, README) from a description |
| HyperIDE | Multi-file IDE with project support, syntax highlighting, AI panel, live preview |
| EpisodicMemory | Conversation history indexed in RAG for semantic recall |
| RAG | Document indexing with embeddings (OpenAI, Mistral, NVIDIA, Google) + IndexedDB persistence |
| Governance | Policy engine + mirror guard + 4-tier trust hierarchy + human kill switch + rollback manager |
| Full Autonomy mode | Grant the AI the same access as the user — bypasses approval prompts (hard barriers still apply) |
git clone https://github.com/AFKmoney/nexusOS.git
cd nexusOS
npm install
npm run devOpen http://localhost:3000. Add an API key in Settings → AI Providers (Mistral, OpenAI, NVIDIA NIM, etc.). Start talking to the AI.
Electron mode (native window + Chromium browser + host filesystem access):
npm run electron:devOne command runs the full pre-merge battery — typecheck, unit tests, build (+ optional lint & e2e) — and exits non-zero if any fatal gate fails:
npm run ci # typecheck + unit tests + build
npm run ci:e2e # also boots the production build under headless ChromeSelect gates explicitly with DEVOP_BATTERY, e.g.:
DEVOP_BATTERY=typecheck,unit,build,node scripts/devops-battery.mjs
RUN_E2E=1 DEVOP_BATTERY=all node scripts/devops-battery.mjs| Gate | Command | Required | Notes |
|---|---|---|---|
typecheck |
tsc --noEmit |
fatal | 0 type errors |
unit |
npm test |
fatal | 0 failing tests |
build |
vite build |
fatal | must succeed |
lint |
eslint (auto) | report-only | enabled when the eslint toolchain is installed; strict via DEVOP_BATTERY_STRICT_LINT=1 |
e2e |
headless Chrome smoke | via RUN_E2E=1 |
boots the prod build, opens every app |
The battery is wired into GitHub Actions (.github/workflows/ci.yml) and runs on every push / pull request.
Tell the DAEMON chat:
- "Build me a pomodoro timer app" — generates a full multi-file app, registers it, you can run it
- "Forge a skill that summarizes my desktop files, then run it" — AI writes a JS skill, persists it, executes it in the sandbox
- "Use csv_inspect on my data, then batch_rename those files" — the AI can call 14 pre-forged pack skills out of the box (data, filesystem, dev, web, productivity)
- "Add a goal to organize my downloads folder, then engage autopilot" — AI works through the goal autonomously
- "What did we talk about yesterday?" — episodic memory recall via RAG
- React 19 + TypeScript + Vite + Tailwind CSS
- Zustand for state, IndexedDB for VFS + RAG persistence
- Electron for native mode (Chromium WebContentsView browser, IPC proxy, host fs)
- Wllama for local GGUF inference (runs entirely in-browser)
- isomorphic-git for VFS git operations
- 346 tests passing (unit + AI component audit + OS action audit)
User ──► DAEMON Chat ──► puterService ──► aiGateway (19 providers + failover)
│
├─► Native Function Calling (35 tools)
│ │
▼ ▼
toolForge ◄── OS:: actions (75)
│
┌───────────────────────┼───────────────────────┐
▼ ▼ ▼
VFS (IndexedDB) AppGenerator AgentOrchestrator v2
(multi-file apps) (parallel + deps)
│ │
▼ ▼
CustomAppRunner SkillForge (Web Worker)
(inlined entry) (sandboxed skills)
Governance pipeline wraps everything: policyEngine → mirrorGuard → trustTier → humanOverride → rollbackManager. Hard barriers (system-reset, self-modify-code, kernel-rules) always require explicit approval, even in Full Autonomy mode.
MIT © Philippe-Antoine Robert
NexusOS is open source. Fork it, extend it, make it yours.