Your private AI assistant — across every channel
Self-hosted. Encrypted by default. Connects to WhatsApp, Telegram, Slack, Discord, and more.
Configure everything by talking — no files to edit, no dashboards to learn.
Quick Start · Vision · Architecture · Security · Contributing
Most AI assistants are stateless — they forget who you are between sessions. Most are cloud-only — your conversations live on someone else's server. Most require technical setup — config files, environment variables, deployment scripts.
GenosOS is different:
- Runs on your hardware. Your conversations never leave your machine.
- Encrypted by default. AES-256-GCM encryption at rest. Not optional — mandatory.
- Remembers you. Structured memory with semantic search. Your assistant knows your preferences, your projects, your context.
- Works across channels. One assistant, connected to WhatsApp, Telegram, Slack, Discord, Signal, iMessage, and more.
- Configured by talking. Say "connect my WhatsApp" instead of editing JSON files.
You: "Connect my WhatsApp"
→ QR code appears, you scan it, done.
You: "Only let my contacts message me"
→ Security policies configured automatically.
You: "Create an assistant for my dental clinic"
→ Specialized agent deployed with appointment scheduling
and patient communication — in one conversation.
# Install Bun (if not installed)
curl -fsSL https://bun.sh/install | bash
# Clone and setup
git clone https://github.com/estebanrfp/genos.git
cd genos
pnpm setup
# Start — asks for API keys on first run, then opens browser
bun genosos.mjs gatewayOn first run, you choose from 3 curated providers — Anthropic, OpenAI, or Gemini — and paste your API key. The browser opens automatically — everything else is configured by talking to the agent.
Core messaging:
| Channel | Library | Status |
|---|---|---|
| Baileys | Production | |
| Telegram | grammY | Production |
| Slack | Bolt | Production |
| Discord | discord.js | Production |
| Signal | signal-utils | Production |
| iMessage | macOS native | Production |
Voice:
| Channel | Technology | Status |
|---|---|---|
| Voice Calls | Twilio/Telnyx + OpenAI Realtime | Production |
| Realtime Call | OpenAI Realtime (bidirectional) | Experimental |
| Local TTS | Kokoro (on-device, no cloud) | Production |
Additional: Google Chat, Microsoft Teams, Matrix, Nostr, Twitch, LINE, WebChat
One assistant, all your channels. Messages from any channel land in the same conversation — with full memory and context.
No config files. No dashboards. The agent understands 30 configuration actions, validated by 190 server-side blueprints. Say what you want — the agent handles the how.
Your assistant remembers who you are across sessions. Structured compaction with TOON encoding reduces token usage by ~40% while preserving meaning. Semantic prefetch injects relevant context before every response — automatically.
Tools are filtered by embedding similarity to the user's intent — only relevant tools reach the LLM. Core tools (read, write, exec, bash) always visible. Domain tools (browser, canvas, cron, image, etc.) appear only when semantically relevant. Saves ~2000-3000 tokens per request. Zero extra API calls — reuses the query embedding from memory prefetch.
Three tested, reliable providers with full tool calling support:
| Provider | Default | Boost |
|---|---|---|
| Anthropic | Claude Sonnet 4.6 | Claude Opus 4.6 |
| OpenAI | GPT-5.4 | o3 |
| Gemini | Gemini 2.5 Pro | Gemini 3 Pro |
No configuration paralysis — pick a provider, paste your key, done.
The LLM auto-escalates to the advanced model when you express importance, urgency, or need for excellence. No manual switching — just say what matters and the boost tool activates the best available model mid-conversation. Works transparently with any configured provider.
Create specialist agents for different domains. A dental clinic assistant, a research agent, a customer service bot — they delegate to each other automatically via A2A communication.
12 pre-built templates: dental clinic, law firm, restaurant, online store, real estate, and more. Deploy a specialized assistant in one conversation.
Bidirectional voice calls via Twilio/Telnyx/Plivo + OpenAI Realtime. Local text-to-speech via Kokoro — no audio leaves your machine.
GenosOS was built with the premise that a personal assistant holds sensitive information — security is not optional.
All config and credentials are encrypted with AES-256-GCM (PBKDF2 100K iterations). Once a vault passphrase is set, every write is encrypted automatically. The agent does not know it is encrypting — it is transparent.
GenosOS was audited against 40,000+ publicly exposed AI assistant instances reported in security research. Every attack vector resolved by default — no manual hardening required.
- Gateway binds to loopback only — refuses to start on public interfaces without auth
- DM pairing by default — unknown senders cannot reach the agent
- Channel tool restrictions — messaging channels cannot run shell commands
- Tamper-evident audit log with HMAC checksums
Full hardening in one command:
bun genosos.mjs agent --message "harden security"Enables: macOS Keychain storage, buffer zeroing, SQLite hardening, Spotlight/Time Machine exclusion, vault auto-lock (30 min), rate limiting.
WhatsApp · Telegram · Slack · Discord · Signal · iMessage · Voice
│
┌─────┴─────┐
│ Gateway │ ← single entry point
│ :18789 │
└─────┬─────┘
│
┌─────────────────┼─────────────────┐
│ │ │
Multi-Agent Memory + TOON Tools + Skills
System (A2A) Compaction semantic filtering
- Runtime: Bun (native SQLite, instant startup)
- Language: Pure JavaScript (ES2024+) — no TypeScript, no build step for development
- Tests: 738 suites, 6,140+ tests (Vitest)
- Extensions: Plugin system with 21 bundled extensions
See ARCHITECTURE.md for design decisions, module map, and roadmap.
pnpm setup # install + build + ui
pnpm test:fast # run unit tests
pnpm gateway:watch # dev loop with hot reloadSee CONTRIBUTING.md for guidelines.
| Document | Description |
|---|---|
| ARCHITECTURE.md | Design decisions, module map, implementation phases |
| VISION.md | Philosophy and technical direction |
| SECURITY.md | Security model and encryption details |
| SECURITY-AUDIT.md | 10-point vulnerability audit |
| docs/ | Conversational guide and configuration reference |
| Mintlify | Official documentation |
| DeepWiki | AI-powered codebase exploration |
Esteban Fuster Pozzi (@estebanrfp) — Full Stack JavaScript Developer
GenosOS is built on the channel infrastructure pioneered by OpenClaw, created by Peter Steinberger and the community. The architecture, security model, memory system, and conversational configuration are original work.