🦀 ClawOS

Your laptop. Your AI. Your rules.

The JARVIS-style AI agent that runs 100% on your hardware —
voice activation, multi-step tool use, 7-layer memory, approval gates, zero cloud.

Install in 2 minutes · See it in action · Read the docs · Join the discussion

---

Why ClawOS?

Ollama gave you a local model. ClawOS gives you a local agent.

ChatGPT runs in the cloud. Jan is a chat wrapper. OpenWebUI is a UI layer.
ClawOS is a full agent OS — it talks, it does, it remembers, and it asks before it acts.

	ClawOS	Odysseus	Open WebUI	Jan
Voice activation	✅ "Hey Claw"	❌	❌	❌
Approval gates	✅ Native popup	❌ Shell access	❌	❌
Agent workflows	✅ 29 built-in	✅ Skills	❌	❌
Multi-agent (A2A)	✅ Agent mesh	❌ Single agent	❌	❌
7-layer memory	✅ Knowledge graph	⚠️ Flat vectors	⚠️ Flat vectors	⚠️ Flat
Bootable ISO	✅ Flash & go	❌	❌	❌
Model manager	✅ Cookbook	✅ Cookbook	✅	✅
Deep research	✅ Citation tracking	✅	❌	❌
Model compare	✅ Side-by-side	✅ Compare	❌	❌
Notes	✅ Markdown + tags	✅	❌	❌
Calendar	✅ iCal export	✅	❌	❌
Email	✅ IMAP + SMTP	✅	❌	❌
PWA / Mobile	✅ Installable	✅	❌	⚠️
Self-hosted	✅ Zero cloud	✅	✅	✅
Zero telemetry	✅ Verify it	✅	✅	✅

---

✨ See it in action

🗣️ Voice Activation — "Hey Claw, good morning"

Say the wake word and hear a synthesized briefing: time, weather, calendar, reminders, what you worked on yesterday. Five tool calls fire in parallel. Fully offline.

You: "Hey Claw, good morning"
Claw: 🔊 "Good morning. It's Tuesday, June 3rd, 72°F and clear.
       You have 3 meetings today, starting with standup at 10.
       Yesterday you pushed 4 commits to ClawOS and closed issue #66.
       Reminder: dentist appointment Thursday."

⚡ Multi-Step Workflows

"Organize my downloads" — 6-tool chain, zero intervention:

1. Scan downloads folder
2. Classify files by type
3. Create category folders
4. Move files
5. Generate summary
6. Report what was done

clawctl run organize-downloads

🛡️ Approval Gates — It asks before it acts

When the agent wants to run a shell command, delete a file, or close an app — a native floating popup appears. You approve or deny. Every time.

This is the #1 differentiator. Other local AI tools give shell access with no guard. ClawOS respects your authority.

🧠 7-Layer Memory — It actually remembers you

Not a goldfish chatbot. ClawOS persists across sessions with structured intelligence:

Layer	What	Example
Pinned facts	Things it should always know	"I prefer dark mode"
Semantic recall	Vector search over conversations	ChromaDB + fastembed
Full-text search	Keyword search across all history	FTS5
Knowledge graph	Entities and relations	"Abrar → works_on → ClawOS"
Archive	Old conversations, compressed	Time-decay compression
ACE learnings	Self-improving corrections	"When I said X, user corrected to Y"
Workflow state	Active task progress	Multi-step task context

---

🚀 Quick start

curl -fsSL https://raw.githubusercontent.com/xbrxr03/clawos/main/install.sh | bash

The installer takes ~2 minutes. Walks you through a 9-step browser wizard — hardware detection, model pull, voice setup, permissions. When done, your dashboard opens at http://localhost:7070.

clawctl health    # verify everything's running
clawctl start     # start all services
clawctl logs      # tail service output

One-command manual install

git clone https://github.com/xbrxr03/clawos.git
cd clawos
pip install -e ".[dev]"
clawctl bootstrap  # interactive setup wizard

Bootable ISO

Got an old laptop? Flash ClawOS onto it and dedicate it to being your JARVIS:

sudo dd if=clawos-amd64.iso of=/dev/sdX bs=4M status=progress

---

⚡ Features

🗣️ Voice — Talk to your computer

• Wake word: "Hey Claw" activates listening — no button needed
• Push-to-talk: Hold a key, speak, release
• Whisper STT: Local speech-to-text, zero cloud
• Piper TTS: Local text-to-speech, natural-sounding
• Morning briefing: Wake up to a voiced summary of your day

clawctl demos morning-briefing    # try it now

🤖 Agent — It does things, not just chats

• Native function calling: qwen2.5 with Ollama-native tool use
• Dynamic model routing: 3b for quick tasks → 7b for reasoning → coder for code
• 31 built-in tools: Shell, files, web search, calendar, reminders, clipboard, screenshot...
• 29 built-in workflows: Organize downloads, summarize PDFs, bulk rename, daily digests...
• Agent mesh (A2A): Multiple agents coordinate on complex tasks

clawctl run organize-downloads    # built-in workflow
clawctl submit "Research Rust vs Go"   # agent task

🛡️ Security — It asks before it acts

• Policy engine: Every tool call goes through policyd
• Approval popup: Sensitive actions trigger a native Tauri window — approve or deny
• Workspace sandbox: File ops can't escape ~/clawos/workspace/
• Shell allowlist: Only approved binaries; blocks python3 -c <code> injection
• Merkle audit trail: Tamper-proof execution log
• No SSRF: Web search blocks private/loopback IPs

Every action is logged. Every sensitive action requires your approval. You are always in control.

🧠 Memory — 7 layers of remembering

• Pinned facts — permanent knowledge ("I prefer dark mode")
• Semantic recall — ChromaDB + fastembed vector search
• Full-text search — FTS5 keyword search across all history
• Knowledge graph — entities and relations via braind
• Archive — compressed old conversations
• ACE learnings — self-improving corrections from feedback
• Workflow state — multi-step task context persistence

clawctl memory search "what did I work on last week"
clawctl memory pin "I prefer dark mode"

🎨 Dashboard — Full control panel

• React SPA at http://localhost:7070
• Workflows: Browse, configure, and run 29 built-in workflows
• Packs: Install curated skill packs
• Traces: Watch agent reasoning in real-time
• Brain: Inspect knowledge graph, search memory
• Settings: Models, voice, permissions, auth — all in one place
• Mobile-responsive: Works on phone and tablet (PWA coming)

📖 Cookbook — Hardware-aware model recommendations

• Auto-detect: Scans CPU, RAM, GPU vendor/VRAM/compute capability
• Smart scoring: 25 models ranked for your exact hardware
• One-command serve: clawctl cookbook serve — picks the best model, pulls, and starts it
• Tier system: A (8GB), B (16GB), C (32GB+) — never recommend what won't fit

clawctl cookbook scan          # detect your hardware
clawctl cookbook recommend    # top 10 models for your rig
clawctl cookbook serve         # auto-pick + start the best model

🔬 Deep Research — Multi-source research with citations

• Search providers: Brave API, Tavily API, or direct URL fetch
• Citation tracking: Sources ranked primary/supporting/tangential
• Session persistence: Research sessions saved to disk, resumable
• Agent integration: Build research intent for agentd synthesis

clawctl research start "quantum computing applications 2026"
clawctl research list
clawctl research get <session-id>

⚖️ Compare — Side-by-side model evaluation

• Parallel execution: Ask multiple models the same question simultaneously
• Per-model metrics: Tokens/sec, total tokens, response time
• Auto-detect running models: No config needed if Ollama is running

clawctl compare "Explain attention mechanisms" --models llama3:8b,qwen2.5:7b

📝 Notes + Calendar + Email — Productivity suite

• Notes: Markdown with YAML front matter, tags, full-text search
• Calendar: Event management, date range filtering, iCal export for external calendars
• Email: IMAP inbox reader + SMTP sender, works with your existing account
• PWA: Install ClawOS on your phone — offline support, push to home screen

clawctl research start "..."   # deep research
# Notes, Calendar, and Email available via dashboard at localhost:7070

🔌 Bring your own brain

• Nexus (built-in) — default agent runtime, optimized for local
• OpenClaw — drop-in power-user agent framework
• Any MCP-compatible — plugin hooks for custom brains
• Ollama — local model serving (qwen2.5, llama3, mistral, phi3...)
• OpenAI-compatible APIs — cloud fallback if you want it (optional)

---

💻 Hardware tiers

Tier	RAM	Model	Experience
A	8 GB	qwen2.5:3b	Basic — works on old laptops, mini PCs
B	16 GB	qwen2.5:7b	Full — multi-step tool use, fast briefings
C	32 GB+	qwen2.5:7b + qwen2.5-coder:7b	Power — coder model for file/shell tasks

Minimum: x86_64 CPU, 8 GB RAM, 20 GB storage, Linux (Ubuntu 22.04+, Fedora 39+, Arch).
NVIDIA GPU optional but accelerates inference.
macOS support arriving in v0.2.

---

🏗️ Architecture

📊 Interactive Mermaid diagram — renders natively on GitHub.

┌──────────────────────────────────────────────────────────┐
│                   User Interfaces                         │
│       Voice  │  Web Dashboard  │  CLI  │  Tauri Overlay  │
└─────────────────────────┬─────────────────────────────────┘
                          │
       ┌──────────────────┼──────────────────┐
       │                  │                  │
   ┌───▼────┐    ┌────────▼─────────┐   ┌────▼─────┐
   │ voiced │    │   Nexus Agent    │   │ Services │
   │  STT   │    │ (4-tier pipeline)│   │  memd    │
   │  TTS   │    │  ┌────────────┐  │   │  policyd │
   │  wake  │    │  │  Policy    │  │   │  workfd  │
   └───┬────┘    │  │  Engine    │  │   │  skilld  │
       │        │  └────────────┘  │   │  desktopd│
       │        └──────────────────┘   └──────────┘
       │
   ┌───▼─────────────────────────────────────────────┐
   │              Local Models (Ollama)               │
   │  qwen2.5:3b · qwen2.5:7b · qwen2.5-coder:7b     │
   └──────────────────────────────────────────────────┘

Core daemons (10 critical at runtime, 29 total):

Daemon	Port	What it does
dashd	7070	Dashboard API + SPA
clawd	7071	Core orchestrator
agentd	7072	Agent runtime + task queue
memd	7073	7-layer memory
policyd	7074	Approval engine + audit log
modeld	7075	Model lifecycle management
voiced	7079	Whisper + Piper voice pipeline
desktopd	7080	Input automation (clipboard, paste, screenshot)
braind	7082	Knowledge graph engine
a2ad	7083	Agent-to-agent mesh protocol

Full service list (29 daemons)

Daemon	Port	What it does
dashd	7070	Dashboard API + SPA
clawd	7071	Core orchestrator
agentd	7072	Agent runtime + task queue
memd	7073	7-layer memory
policyd	7074	Approval engine + audit log
modeld	7075	Model lifecycle management
metricd	7076	Metrics collection
mcpd	7077	MCP tool server
observd	7078	Observability
voiced	7079	Whisper + Piper voice pipeline
desktopd	7080	Input automation
agentd_v2	7081	Next-gen agent runtime
braind	7082	Knowledge graph
a2ad	7083	Agent-to-agent mesh
sandboxd	7085	Sandbox execution
visuald	7086	Visual processing
reminderd	7087	Desktop notifications
waketrd	7088	Wake word → briefing bridge

---

🔒 Zero telemetry — verify it yourself

No analytics. No error reporting. No usage stats. No phone-home. Run this:

grep -rn "posthog\|sentry\|google-analytics\|mixpanel\|amplitude" \
  --include="*.py" --include="*.ts" --include="*.tsx" .
# returns nothing

The only network calls ClawOS makes: Ollama (localhost), DuckDuckGo (web search), wttr.in (weather), your RSS feeds. All optional. All disableable.

---

🛠️ Development

git clone https://github.com/xbrxr03/clawos.git
cd clawos
pip install -e ".[dev]"

# Run tests (no live LLM needed)
pytest tests/ -q

# Boot dev services
bash scripts/dev_boot.sh --full

# Check health
clawctl health

# Tail logs
clawctl logs dashd

See CONTRIBUTING.md for guidelines.

---

📁 Project structure

clawos/
├── runtimes/agent/         # Nexus agent loop (4-tier pipeline)
│   ├── runtime.py          #   Priority pipeline: memory → confirm → intent → LLM
│   ├── intents.py          #   Deterministic regex classifier
│   ├── router.py           #   3b/7b/coder dynamic model router
│   ├── tool_schemas.py     #   31 tool JSON schemas for Ollama function calling
│   ├── briefing.py         #   Morning briefing
│   └── tools/              #   8 tool modules (Linux + macOS)
├── services/               # 29 daemons (FastAPI + SQLite)
├── workflows/              # 28 built-in workflows
├── desktop/command-center/ # Tauri shell + approval overlay
├── dashboard/frontend/     # React SPA
├── clawctl/                # CLI
├── packaging/              # AppImage, .deb, AUR, ISO
└── tests/                  # 479 unit + integration tests

---

📚 Documentation

Doc	What's inside
Demos walk-through	Exact phrasing and expected output for each demo
Architecture overview	How the pieces fit together
Architecture diagram	Mermaid diagram of system and request flow
CLI reference	All clawctl commands, flags, and examples
API reference	Service endpoints and contracts
Security audit	Threat model and mitigations
Product vision	Where we're headed
Roadmap	Milestones and current status

---

⭐ Star this repo

If you want local AI that actually does things — not just chat — star ClawOS and follow the progress.

Every star tells us: build this faster.

⭐ Star = "I want this on my machine"

---

🤝 Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

git checkout -b feature/my-feature
git commit -m "feat: add awesome feature"
git push origin feature/my-feature
# Open a PR

---

📜 License

This project is licensed under the GNU Affero General Public License v3 or later (AGPL-3.0-or-later). Forks must remain open source.

---

🙏 Acknowledgments

• Ollama — local LLM serving
• Qwen team — qwen2.5 models
• Piper — local TTS
• OpenClaw — optional power-user agent brain

---

The future of AI is local, private, and yours to control.

Get started · Read the docs · Join the discussion