Michael Limberger

The mixer that did not happen

This talk wires OpenClaw to Ollama on Apple Silicon. The assistant stays on localhost with models on disk.

The research was already done and the notes were written, but the live meetup was cancelled over security concerns. The topic is still useful if you understand the risks and make your own call. So here it is, free, with the hardening section treated as required reading rather than a footnote.

What this stack is

OpenClaw is an open source personal AI agent by Peter Steinberger. It can connect to WhatsApp, Telegram, Discord, and other platforms while running on your local machine. Pair it with Ollama and a 14 billion parameter language model, and you get a private assistant that keeps inference on your hardware instead of a cloud chat vendor.

You pay with your own machine, not a subscription. The model and the agent stay under your rules, as long as you install and configure them carefully.

Why local, why now

Agent ecosystems on the public internet move fast, and the news cycle around them is loud. Running your stack locally, with models on disk and no need to ship chat to a third party, is ordinary operational hygiene for anyone who cares about credentials, mail, and private notes.

That does not mean OpenClaw itself is risk-free. The security section covers why this guide is detailed about lockdown.

The walk tonight

We go end to end in ten parts.

1. This overview. Why we're here, what the risks are, what you're
           getting into.
        
        2. Hardware requirements. What your Mac actually needs to run a 14B
           model without melting.
        
        3. Installing Ollama. The local model server that handles GPU
           acceleration on Apple Silicon.
        
        4. Choosing and pulling a model. Not all 14B models are equal, and
           the wrong choice will waste your time.
        
        5. Installing OpenClaw. Three methods, each with their own gotchas.
        
        6. Configuration. The JSON config file that tells OpenClaw to use
           Ollama instead of cloud providers.
        
        7. Security hardening. This is the big one. The ClawHavoc supply
           chain attack compromised 824 skills in early 2026. We cover
           everything you need to lock down.
        
        8. Using OpenClaw. Dashboard, terminal UI, command line. No
           messaging platforms required.
        
        9. Troubleshooting. Every common error, why it happens, how to
           fix it.
        
        10. Performance tuning and testing. Getting the most out of Apple
            Silicon's unified memory, and verifying the whole stack works.

+------------------------------------------------------------------+
|  USER DISCRETION ADVISED                                         |
|                                                                  |
|  This guide documents how to install third party software that   |
|  has had serious security incidents. Read the security section   |
|  before installing anything. Make your own risk assessment.      |
|  The author cancelled the live event over these concerns.        |
+------------------------------------------------------------------+

ClawHavoc, in brief

In late January 2026, a coordinated supply chain attack called ClawHavoc hit the ClawHub skill marketplace. At its peak, 824 of approximately 10,700 skills (about 8%) were trojanized. The payload was the Atomic macOS Stealer (AMOS), which harvests browser credentials, SSH keys, Telegram sessions, and cryptocurrency wallets.

Separately, security researchers found over 312,000 OpenClaw instances exposed on the public internet, many without authentication. Five CVEs were disclosed, including CVE-2026-25253, a CVSS 8.8 one-click remote code execution through the Control UI.

This is recent history, not folklore. If you follow this tutorial, treat the security hardening section as mandatory. Skip it only if you have already decided the risk is yours to carry.

What you need

Hardware:

Apple Silicon Mac (M1, M2, M3, or M4)
        Minimum 16GB unified memory (24GB or higher strongly recommended)
        At least 15GB free disk space
        macOS Sonoma (14) or newer

Intel Macs and 8GB machines are out of scope here. Check Apple menu, About This Mac, if you are unsure.

Software (installed in later parts):

Ollama (local model server)
        OpenClaw (the AI agent framework)
        Node.js 22.12+ (OpenClaw dependency, auto-installed by the installer)

Time: about 15 minutes if the model is already downloaded, or about 30 minutes if you pull it live on a decent connection.

Next up: hardware requirements in more detail.