CLAWOS ONLINE  ·  Open Source · AGPL-3.0

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.

curl -fsSL https://clawos.ai/install.sh | bash
⭐ Star on GitHub Install in 2 min →
479
Tests passing
29
Workflows
0
Cloud calls
7
Memory layers
Why ClawOS
Ollama gave you a local model.
ClawOS gives you a local agent.

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

Feature 🦀 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
Structured memory ✅ 7 layers ⚠️ Flat vectors ⚠️ Flat vectors ⚠️ Flat
Bootable ISO ✅ Flash & go
Model manager 🚧 Cookbook
Deep research 🚧 Phase 3
Zero telemetry ✅ Verify it
What it does
An agent, not a chatbot.

ClawOS runs 29 multi-step workflows, controls 31 tools, coordinates multiple agents, and remembers everything across sessions.

🗣️
Voice Activation
"Hey Claw" wake word, Whisper STT, Piper TTS. Morning briefings voiced. Fully offline.
WhisperPiperVAD
Multi-Step Workflows
29 built-in automations: organize downloads, summarize PDFs, bulk rename, daily digests. One command.
29 workflows6 categories
🛡️
Approval Gates
Native floating popup when the agent wants to run a shell command, delete a file, or close an app. You approve. Every time.
TauripolicydMerkle audit
🧠
7-Layer Memory
Pinned facts, semantic recall, FTS5 search, knowledge graph, archive, ACE learnings, workflow state. It actually remembers.
ChromaDBfastembedbraindFTS5
🤖
Agent Mesh (A2A)
Multiple agents coordinate on complex tasks. Peer discovery, inter-agent communication, per-agent policy gates.
a2adagentdpolicyd
🎨
Dashboard
React SPA with workflows, packs, traces, brain viewer, and settings. Mobile responsive. PWA coming.
ReactFastAPIWebSocket
Talk to your computer
"Hey Claw, good morning"

Say the wake word. Five tool calls fire in parallel. You hear a synthesized briefing — time, weather, calendar, what you did yesterday. Offline.

Live terminal session
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." # Behind the scenes: 5 parallel tool calls → weather.current() → calendar.today() → git.recent() → memory.search("yesterday") → reminders.upcoming()
One-command workflows
$ clawctl run organize-downloads → Scanning ~/Downloads ... 47 files found → Classifying: 12 documents, 8 images, 15 archives, 12 other → Creating folders: Documents/, Images/, Archives/ → Moving 35 files to categories ✓ Organized 47 files into 4 categories in 2.3s
It actually remembers you
7 layers of memory.

Not a goldfish chatbot. ClawOS persists across sessions with structured intelligence — facts, semantics, graphs, corrections, and compressed archives.

Layer 1
Pinned Facts
Things it should always know — "I prefer dark mode", "My timezone is ET"
SQLite
Layer 2
Semantic Recall
Vector search over all conversations — find what you meant, not just what you said
ChromaDB + fastembed
Layer 3
Full-Text Search
Keyword search across every conversation you've ever had
SQLite FTS5
Layer 4
Knowledge Graph
Entities and relations — "Abrar works on ClawOS" — structured intelligence
braind
Layer 5
Archive
Old conversations, compressed with time-decay. Accessible but not in your way
gzip + time-decay
Layer 6
ACE Learnings
Self-improving — "When I said X, user corrected to Y" — gets smarter over time
correction log
Layer 7
Workflow State
Active multi-step task context — pick up where you left off
taskd
Security first
It asks before it acts.

Every sensitive action requires your approval. Every action is logged. You are always in control. This is the #1 differentiator — other local AI tools give shell access with no guard.

runtime
Approval Gates
Native floating popup for shell commands, file deletions, app control. Approve or deny.
policy
Policy Engine
Every tool call goes through policyd. Per-agent permission profiles.
sandbox
Workspace Sandbox
File ops can't escape ~/clawos/workspace/. Sandboxed by default.
audit
Merkle Audit Trail
Tamper-proof execution log. Every action, every approval, every denial.
shell
Shell Allowlist
Only approved binaries. Blocks python3 -c injection, pipe bombs, and privilege escalation.
network
No SSRF
Web search blocks private/loopback IPs. Your internal network stays internal.
Runs on what you have
Three hardware tiers.

From an old mini PC to a beefy workstation — ClawOS adapts to your hardware.

A
Basic
8 GB RAM
Any x86_64 CPU
qwen2.5:3b
B
Full
16 GB RAM
Multi-core CPU
qwen2.5:7b
C
Power
32 GB+ RAM
NVIDIA GPU optional
7b + coder:7b
Get started
Install in 2 minutes.

One command. 9-step browser wizard handles hardware detection, model pull, voice setup, and permissions.

Quick install
$ curl -fsSL https://clawos.ai/install.sh | bash → Detecting hardware ... x86_64, 16GB RAM, RTX 3060 → Pulling recommended model ... qwen2.5:7b (4.7 GB) → Configuring voice ... Whisper STT + Piper TTS ✓ → Starting services ... 10/10 ✓ ✓ ClawOS running at http://localhost:7070 $ clawctl health dashd ✓ clawd ✓ agentd ✓ memd ✓ policyd ✓ modeld ✓ voiced ✓ braind ✓ a2ad ✓ workfd ✓
Manual install
$ git clone https://github.com/xbrxr03/clawos.git $ cd clawos && pip install -e ".[dev]" $ clawctl bootstrap # interactive setup wizard

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

Join the community. Star the repo. Build with us.