clawspark

One command. Private AI agent. Your hardware.

Website · Tutorial · Configuration · Contributing

---

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

That's it. Come back in 5 minutes to a fully working, fully private AI agent that can code, research, browse the web, analyze images, and manage your tasks. Everything runs on your hardware. No cloud APIs, no subscriptions, no telemetry.

What is this?

OpenClaw is the most popular open-source AI agent (340K+ stars). clawspark gets it running on your NVIDIA hardware in one command. Fully local. Fully private. Your data never leaves your machine.

What happens when you run it:

1. Detects your hardware (DGX Spark, Jetson, RTX GPUs, Mac)
2. Picks the best model using llmfit for hardware-aware selection
3. Installs everything (Ollama, OpenClaw, 10 skills, dependencies)
4. Configures multi-model (chat + vision + optional image generation)
5. Enables voice (local Whisper transcription, zero cloud)
6. Sets up browser automation (headless Chromium)
7. Sets up your dashboard (chat UI + metrics)
8. Creates systemd services (auto-starts on boot)
9. Hardens security (firewall, auth tokens, localhost binding, Docker sandbox)

Supported Hardware

Hardware	Memory	Default Model	Tokens/sec
DGX Spark	128 GB unified	Qwen 3.5 35B-A3B	~59 (measured)
Jetson AGX Thor	128 GB unified	Auto-selected	Community testing
Jetson AGX Orin	64 GB unified	Auto-selected	Community testing
RTX 5090 / 4090	24-32 GB VRAM	Auto-selected	Community testing
RTX 4080 / 4070	8-16 GB VRAM	Auto-selected	Community testing
Mac M1/M2/M3/M4	16-128 GB unified	Auto-selected	Community testing

NVIDIA platforms use llmfit to detect your hardware and pick the best model. macOS uses a curated fallback list.

Quick Start

The installer asks 3 questions:

[1/3] Which model?         > 5 models ranked by hardware fit
[2/3] Messaging platform?  > WhatsApp / Telegram / Both / Skip
[3/3] Tailscale?           > Yes (remote access) / No

Zero interaction mode:

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

What Your Agent Can Do

Capability	How it Works
Answer questions	Local LLM via Ollama
Search the web	Built-in web search + DuckDuckGo, no API key
Deep research	Sub-agents run parallel research threads
Browse websites	Headless Chromium (navigate, click, fill forms, screenshot)
Analyze images	Vision model for screenshots, photos, diagrams
Write and run code	exec + read/write/edit tools
Voice notes	Local Whisper transcription for WhatsApp voice messages
File management	Read, write, edit, search files on the host
Scheduled tasks	Cron-based automation
Sub-agent orchestration	Spawn parallel background agents

All of this runs locally. No data leaves your machine.

Skills

10 verified skills ship by default. Install curated bundles:

clawspark skills pack research      # Deep research + web search (4 skills)
clawspark skills pack coding        # Code generation + review (2 skills)
clawspark skills pack productivity  # Task management + knowledge (3 skills)
clawspark skills pack voice         # Voice interaction (2 skills)
clawspark skills pack full          # Everything (10 skills)

Manage individual skills:

clawspark skills add <name>         # Install a skill
clawspark skills remove <name>      # Remove a skill
clawspark skills sync               # Apply skills.yaml changes
clawspark skills audit              # Security scan installed skills

Multi-Model

Three model slots:

Slot	Purpose	Example
Chat	Conversation and coding	ollama/qwen3.5:35b-a3b
Vision	Image analysis	ollama/qwen2.5-vl:7b
Image gen	Create images (optional)	Local ComfyUI or API

clawspark model list                # Show all models
clawspark model switch <model>      # Change chat model
clawspark model vision <model>      # Set vision model

Security

• UFW firewall (deny incoming by default)
• 256-bit auth token for the gateway API
• Gateway binds to localhost only
• Code-level tool restrictions (21 blocked command patterns)
• SOUL.md + TOOLS.md with immutable guardrails
• Plugin approval hooks (user confirmation before acting)
• Optional Docker sandbox (no network, read-only root, all caps dropped)
• Air-gap mode: clawspark airgap on
• OpenAI-compatible API gateway for local-first workflows

Skill security audit -- scans installed skills for 30+ malicious patterns (credential theft, exfiltration, obfuscation). Protects against ClawHub supply chain attacks:

clawspark skills audit

Diagnostics

Full system health check across hardware, GPU, Ollama, OpenClaw, skills, ports, security, and logs:

clawspark diagnose                  # alias: clawspark doctor

Generates a shareable debug report at ~/.clawspark/diagnose-report.txt.

CLI Reference

clawspark status              Show system health
clawspark start               Start all services
clawspark stop [--all]        Stop services (--all includes Ollama)
clawspark restart             Restart everything
clawspark update              Update OpenClaw, re-apply patches
clawspark benchmark           Run performance benchmark
clawspark model list|switch|vision   Manage models
clawspark skills sync|add|remove|pack|audit   Manage skills
clawspark sandbox on|off|status|test   Docker sandbox
clawspark tools list|enable|disable   Agent tools
clawspark mcp list|setup|add|remove   MCP servers
clawspark tailscale setup|status   Remote access
clawspark airgap on|off       Network isolation
clawspark diagnose            System diagnostics
clawspark logs                Tail gateway logs
clawspark uninstall           Remove everything

Dashboard

Two web interfaces out of the box:

• Chat UI: http://localhost:18789/__openclaw__/canvas/
• Metrics: http://localhost:8900 (ClawMetry)

Both bind to localhost. Use Tailscale for remote access.

Docker Sandbox

Optional isolated code execution for sub-agents:

clawspark sandbox on          # Enable
clawspark sandbox off         # Disable
clawspark sandbox test        # Verify isolation

Containers run with no network, read-only root, all capabilities dropped, custom seccomp profile, and memory/CPU limits.

Uninstall

clawspark uninstall

Removes all services, models, and config. Conversations preserved in ~/.openclaw/backups/ unless you pass --purge.

Testing

73 tests using bats:

bash tests/run.sh

Suite	Tests	Coverage
common.bats	27	Logging, colors, helpers
skills.bats	16	YAML parsing, add/remove, packs
security.bats	11	Token generation, permissions, deny lists
cli.bats	19	Version, help, routing, error handling

Acknowledgements

• OpenClaw -- AI agent framework
• Ollama -- Local LLM inference
• llmfit -- Hardware-aware model selection
• Baileys -- WhatsApp Web client
• Whisper -- Speech-to-text
• ClawMetry -- Observability dashboard
• Qwen -- The model family that runs great on DGX Spark

Maintainers

• Saiyam Pathak
• Rohit Ghumare

Contributing

PRs welcome. Areas where help is needed:

• Testing on Jetson variants and RTX GPUs
• Hardware detection for more GPU models
• Additional messaging platform integrations
• New skills and skill packs
• Sandbox improvements

License

MIT. See LICENSE.

---

Built for people who want AI that works for them, not the other way around.
clawspark.dev