Self-hosted AI agent runtime — your personal assistant that lives on your hardware.
The reference implementation of the AGEX protocol for agent identity and trust.
QuantumClaw is an AI agent that runs on your machine — laptop, VPS, Raspberry Pi, or Android phone. It connects to your messaging apps, learns about you over time, and takes actions on your behalf.
- Runs where your data lives — no cloud, no tracking, your conversations never leave your hardware
- Talks to you everywhere — Telegram, Discord, WhatsApp, Slack, Email, and a web dashboard
- Gets smarter over time — extracts facts, preferences, and events from every conversation
- Acts autonomously — runs commands, reads/writes files, searches the web, sends emails, all governed by your safety rules
- Speaks and listens — send a voice note on Telegram, get a voice reply back
npm i -g quantumclaw
qclaw onboard # interactive setup
qclaw start # agent + dashboard at localhost:3000Requirements: Node.js 20+ and one LLM API key (Anthropic, OpenAI, Groq free tier, OpenRouter, Google, xAI, Mistral, or Together).
Everything local. SQLite database, AES-256-GCM encrypted secrets, direct LLM API calls — no middleware.
| Channel | Features |
|---|---|
| Telegram | DMs, voice transcription + TTS replies, image support |
| Discord | @mentions in servers, DMs, 2000-char splitting |
| QR code pairing, session persistence, group filtering | |
| Slack | Socket Mode, @mentions + DMs, 4000-char splitting |
| IMAP polling + SMTP auto-reply |
All channels use a pairing flow — your agent only talks to people you approve. Assign different agents to different channels via config.
Three-layer memory that survives restarts: vector search (TF-IDF), structured knowledge store (facts, events, preferences), and optional Cognee-powered knowledge graph. Per-agent isolation. Dashboard has search, remember/forget, graph visualization, and JSON export.
Spawn specialist agents from CLI, dashboard, or conversation. Each agent has its own SOUL.md personality, AGEX cryptographic identity, scoped tool access, and conversation history.
shell_exec, read_file, write_file, list_directory, web_fetch, render_canvas, search_knowledge, spawn_agent, get_current_time, calculate — plus 12 pre-configured MCP servers and 3,286 community skills from ClawHub.
Telegram voice notes are transcribed (Deepgram → Whisper → Groq) and replies come back as voice (ElevenLabs → OpenAI TTS). Images work via paste/drag-drop in dashboard.
Agent renders HTML, SVG, Mermaid diagrams, and Markdown directly in a split-pane dashboard view. Multiple artifacts persist as tabs.
Agent sends messages to all your channels unprompted — morning briefs, price alerts, weekly summaries. Configure via dashboard or config.
Immutable safety rules in VALUES.md. Every tool call is checked before execution. The agent cannot modify its own rules.
Automatic model selection: reflex (free) → simple → standard → complex → expert. Average daily cost: £0.01–0.05.
12-page web UI at http://localhost:3000:
💬 Chat (with Live Canvas) · 📊 Overview · 📡 Channels · 📈 Usage · 🤖 Agents · 🔧 Skills · ⚡ Tools · ⏰ Tasks · 🧠 Memory · 🔑 API Keys · ⚙️ Config · 📋 Logs
qclaw start / stop / restart / status
qclaw chat # terminal chat
qclaw dashboard # show URL
qclaw agent list / spawn / delete
qclaw skill list / search / install / remove
qclaw tool list / enable / disable
qclaw secret set / list
qclaw pairing list / approve
qclaw config get / set
qclaw updatecp .env.example .env && docker compose up -dChannels (5) → 5-Tier Router → Agent Registry → Tools (10+ built-in, MCP, ClawHub)
↓ ↓
Trust Kernel Memory (3 layers)
(VALUES.md) Vector · Knowledge · Graph
AGEX protocol gives every agent a cryptographic identity for authentication, scoped permissions, and credential sharing.
MIT © 2025–2026 ALLIN1.APP LTD — Built by Hayley