Open Source / MIT License

clawspark

One command. Private AI agent. Your hardware.

Turn your DGX Spark, RTX GPU, Mac, or Jetson into a private AI agent that can code, research, browse the web, analyze images, and run tasks. Everything runs locally. Your data never leaves your machine.

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

What Your Agent Can Do

Not just a chatbot. A full AI agent that writes code, browses the web, analyzes images, and runs tasks on your hardware.

Write and Run Code

Ask it to build apps, scripts, or tools. It writes the code, tests it in a Docker sandbox, and iterates until it works.

Deep Research

Spawns parallel sub-agents that each research a subtopic, then merges everything into one report. Web search via DuckDuckGo, no API key needed.

Browser Automation

Headless Chromium built in. Navigate websites, fill forms, click buttons, take screenshots. All running locally.

Vision and Image Analysis

Send a screenshot, photo, or diagram. The vision model (qwen2.5-vl) analyzes it and describes what it sees. Multi-model support built in.

Voice Enabled

Send WhatsApp or Telegram voice notes. Whisper transcribes them locally on your GPU. No audio ever leaves your machine.

Docker Sandbox

Agent-generated code runs in isolated containers. No network, read-only root, all capabilities dropped, custom seccomp profile.

Sub-Agent Orchestration

Spawns background agents for parallel work. Up to 4 concurrent sub-agents with 2 levels of depth. Complex tasks broken into threads.

Skill Packs

Curated bundles for research, coding, productivity, and voice. Install a full skill pack with one command or pick individual skills.

Diagrams and Architecture

Ask for a flowchart, architecture diagram, or mindmap. MCP-powered Mermaid generates real PNG images from text. 20+ diagram types.

Survives Reboots

Systemd services auto-start everything on boot. Gateway, node host, dashboard, Ollama. No manual restarts, ever.

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.

Docker Sandbox

Code execution isolated in containers: no network, read-only root, all capabilities dropped, custom seccomp profile, 200 PID limit.

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 / Voice
           |
  OpenClaw Gateway  (port 18789)
    |              |            |           |
  Agent          Node Host    Baileys      Sub-Agents
  (LLM)         (28+ tools)  (WhatsApp)   (parallel)
    |              |              |
  Ollama       exec, browser,   MCP Servers
  (port 11434)  read, write,     (mcporter)
    |           web_search,      |
  Chat Model  vision      Mermaid, Memory,
  + Vision        |        Filesystem, Thinking
  (GB10 GPU)  Docker Sandbox
              (isolated exec)
128 GB
Unified Memory
~59 tok/s
Generation Speed
1 PFLOP
AI Compute
28+ Tools
Agent + MCP Tools

Supported Hardware

clawspark automatically detects your hardware and selects the best model configuration.

Hardware Memory Default Model Status
NVIDIA DGX Spark 128 GB unified Qwen 3.5 35B-A3B ~59 tok/s (tested)
Mac M1 / M2 / M3 / M4 16-128 GB unified llmfit-selected untested
RTX 5090 / 4090 24-32 GB VRAM llmfit-selected untested
RTX 4080 / 4070 8-16 GB VRAM llmfit-selected untested
Jetson AGX Orin / Thor 64-128 GB unified llmfit-selected untested

DGX Spark is the only platform with full end-to-end testing. Other platforms should work -- community testing welcome.

How It Works

From bare hardware to a fully capable AI agent in about 5 minutes.

Step 1

Detect Hardware

Scans for GPUs, identifies your hardware model, and measures available VRAM or unified memory.

Step 2

Pick the Best Model

Uses llmfit to select the optimal chat model for your hardware. Pulls a vision model automatically for image analysis.

Step 3

Install Everything

Ollama, OpenClaw, skills, Whisper for voice, Chromium for browser automation. All handled automatically.

Step 4

Set Up Sandbox

Installs Docker if needed, builds an isolated container image with seccomp, no-network, and read-only rootfs for safe code execution.

Step 5

Create Services

Systemd services for gateway, node host, and dashboard. Everything auto-starts on boot and restarts on failure.

Step 6

Harden Security

Firewall rules, auth tokens, localhost binding, read-only guardrails. Your agent is locked down by default.

Done

Agent Ready

Your private AI agent is live. Code, research, browse, analyze images, run tasks. All on your hardware, all private.

Tools Installed

26+ agent tools, 10 skills, and a full management CLI 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 (26+ 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
  • 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 (qwen2.5-vl)
  • browser -- Navigate websites, click, fill forms, screenshot
  • web_search -- DuckDuckGo search, no API key
  • image -- Download and process images
  • pdf -- Read and extract PDF content
  • sessions_spawn -- Sub-agent sessions for parallel work
Scroll for all tools

Management CLI

  • clawspark status -- Check all components
  • clawspark start -- Start all services
  • clawspark stop -- Stop all services
  • clawspark restart -- Restart services
  • clawspark model list -- Show all models
  • clawspark model switch -- Change chat model
  • clawspark sandbox on -- Enable Docker sandbox
  • clawspark skills pack -- Install skill bundles
  • clawspark logs -- View logs

Get Started

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

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

Works on DGX Spark, RTX GPUs (Linux + CUDA), Mac (Apple Silicon), and Jetson. Fully tested on DGX Spark.