Open Source / MIT License

clawspark

One command. Private AI assistant. Your hardware.

The one-click installer for OpenClaw on NVIDIA hardware. Turn your DGX Spark, Jetson, or RTX GPU into a private AI assistant you control via WhatsApp and Telegram, with local voice transcription and full privacy.

$ curl -fsSL https://clawspark.dev/install.sh | bash

What It Does

Everything you need to go from bare hardware to a fully private AI assistant, handled automatically.

Hardware Detection

Auto-detects your NVIDIA hardware: DGX Spark, Jetson, or RTX GPUs. No manual configuration needed.

Best Model Selection

Picks the optimal AI model for your specific hardware, maximizing performance without running out of memory.

One-Click Install

OpenClaw, inference engine, and all skills deployed in a single command. No Docker headaches, no dependency hell.

Voice Enabled

Send WhatsApp or Telegram voice notes. Whisper transcribes them locally on your GPU, keeping conversations private.

Dashboard + Metrics

Built-in chat UI and ClawMetry metrics dashboard. Track tokens, costs, sessions, and agent activity in real time.

Remote Access

Optional Tailscale integration. Access your AI from your phone in another country, securely over your Tailnet.

Security

Your AI runs on your hardware and stays locked down. Every layer is hardened by default.

Localhost Only

Gateway binds to 127.0.0.1. Your AI is not accessible from the network.

Token Authentication

Random 256-bit token generated at install. Every API call requires it.

Tool Restrictions

Dangerous tools (exec, write, browser, process) blocked by default. Bot cannot run commands on your machine.

Hardened Agent

SOUL.md and TOOLS.md are read-only. The AI cannot modify its own guardrails.

Firewall Rules

UFW configured automatically. Deny incoming, allow outgoing, SSH only.

Air Gap Mode

Optional complete network isolation via clawspark airgap on. Zero outbound traffic.

Architecture

A clean stack from messaging layer to GPU. Everything runs locally on a single machine.

  WhatsApp / Telegram / Web UI
           |
  OpenClaw Gateway  (port 18789)
    |              |            |
  Agent          Node Host    Baileys
  (LLM)         (Tools)      (WhatsApp)
    |              |
  Ollama       read, web_fetch,
  (port 11434)  message
    |
  Qwen 3.5 35B
  (GB10 GPU)
128 GB
Unified Memory
~59 tok/s
Generation Speed
1 PFLOP
AI Compute
15 Tools
Agent Capabilities

Supported Hardware

clawspark automatically detects your GPU and selects the best configuration.

Hardware VRAM Recommended Model Performance
NVIDIA DGX Spark 128 GB Qwen 3.5 35B-A3B (MoE) / Qwen 3.5 122B ~59 tok/s
Jetson AGX Orin 64 GB Nemotron 3 Nano 30B ~25 tok/s
RTX 5090 32 GB Qwen 3.5 35B-A3B (Q4) ~40 tok/s
RTX 4090 24 GB GLM 4.7 Flash ~35 tok/s

How It Works

From bare hardware to a working private AI assistant in minutes.

Step 1

Detect

Scans your system for NVIDIA GPUs and identifies the exact hardware model and available VRAM.

Step 2

Recommend

Selects the best AI model, quantization level, and inference engine for your specific hardware.

Step 3

Download

Pulls the model weights and required containers. Progress bars keep you informed.

Step 4

Install

Deploys OpenClaw with the inference engine, Whisper for voice, and all configured skills.

Step 5

Secure

Locks down the installation: localhost-only access, firewall rules, and optional air gap mode.

Step 6

Connect

Guides you through linking WhatsApp or Telegram so you can start messaging your AI assistant.

Done

Ready

Your private AI assistant is live. Talk to it via voice or text, entirely on your own hardware.

Tools Installed

Skills, tools, and CLI commands that ship with every clawspark installation.

Agent Skills

  • Web Search -- DuckDuckGo, no API key needed
  • Deep Research -- Multi-source synthesis
  • Voice Transcription -- Local Whisper on GPU
  • Prompt Guard -- Injection defense

Agent Tools (15 tools, work out of the box)

  • exec -- Shell commands (bash, docker, kubectl)
  • web_fetch -- HTTP requests, API calls, web search via DDG
  • read -- Read files on the host
  • write -- Write and create files
  • edit -- Edit files in place
  • message -- WhatsApp / Telegram replies
  • canvas -- Interactive web UI
  • process -- List, monitor, kill processes
  • cron -- Scheduled tasks
  • sessions_spawn -- Sub-agent sessions for parallel work
  • nodes -- Remote node execution (paired devices)
  • memory_search -- Search stored context
  • memory_store -- Save info across sessions
  • transcribe -- Audio/voice transcription (local Whisper)
  • vision -- Image analysis (model-dependent)
Scroll for all 15 tools

Management CLI

  • clawspark status -- Check all components
  • clawspark restart -- Restart services
  • clawspark update -- Update OpenClaw + re-apply patches
  • clawspark logs -- View logs

Get Started

One command is all it takes. Run it on any supported NVIDIA machine.

$ curl -fsSL https://clawspark.dev/install.sh | bash

Requires an NVIDIA GPU with CUDA drivers installed. Works on Ubuntu 22.04+ and JetPack 6+.