Verdict
For 70B-class models on one machine with no cloud call, a Mac Studio running Ollama and ZeroClaw keeps the whole loop local: ZeroClaw sends its requests to the Ollama server on the same Mac. The agent itself still reaches the network unless you restrict it. The price rules out most hobbyist use. Until October 2026 this page said ZeroClaw denies network egress and has an air-gap mode; its docs describe neither.
Setup notes
macOS host. Ollama with Llama 3.3 70B Q4, then ZeroClaw from its install script, zeroclaw quickstart and Ollama as the provider. ZeroClaw's docs describe no air-gap mode: to keep the machine offline, cut its network at the host and update the binary by hand.
Performance
No published 70B speed: llama.cpp's scoreboard has no 70B row (a 7B model writes 92.14 tokens per second on the 80-core GPU). 96 to 256 GB of unified memory means several models can stay loaded at once.
What breaks
- Cost-sensitive setups
- Linux-only production stacks
Sources
The agent's defaults and settings above were read in its own documentation on 2026-10-18.
- ZeroClaw README (v0.8.6): providers, channels, install, licence
- ZeroClaw NOTICE: dual licence, MIT or Apache-2.0
- ZeroClaw docs (v0.8.6): security model, default posture
- ZeroClaw docs (v0.8.6): sandboxing (auto-detection, network egress)
- ZeroClaw docs (v0.8.6): provider-agnostic
Want to know more
See the ZeroClaw page and the Mac Studio M3 Ultra spec sheet.