Omarchy Local AI closes a shortcut around model gateway keys
Version 6.17.0 puts each serving engine on a private Docker network after maintainers identified direct access that could bypass the keyed gateway.
A password-protected model gateway is little help if another local account can reach the engine behind it. Local AI for Omarchy’s October 15 version 6.17.0 addresses that boundary: maintainers say each model engine now runs on an internal Docker network shared only with its gateway, with no host address on that network.
The release explains that Docker previously routed bridge addresses from the host even without a published port. A local account could therefore reach an engine directly and bypass the gateway key, particularly when TabbyAPI used disable_auth or llama-server ran without --api-key. This is a documented access path, not evidence of exploitation.
The upgrade changes the model’s network
According to the release, the gateway joins both the private engine network and the bridge used by Local AI and its agents. Existing models move to their private network on the first panel refresh without restarting. Engines also lose outbound network access; weights are mounted and Hub downloads are already offline.
Docker 28 or later is required, and an older Docker refuses model startup with an explanation. The release also adds Intel Arc Pro B65 detection and three llama.cpp SYCL recipes. AITrending has not independently tested the isolation or those hardware recipes.
A desktop plugin for existing agents
The documented installation is omarchy plugin add https://github.com/sybil-solutions/omarchy-local-ai --enable. It requires Omarchy with the Quattro shell and bar. After downloading a compatible model, users choose a folder and open an existing harness such as pi, Claude Code, Codex or OpenCode against the local gateway; the plugin says it leaves each harness’s own configuration untouched.
Setup asks for permission and a password to enable Omarchy’s root-equivalent Sudoless Docker access, and installs NVIDIA container support when needed. A Hugging Face token is optional for downloads when present. Local inference consumes the user’s hardware rather than a hosted inference subscription; model licenses and optional external services still have separate terms. Usage reporting can be disabled in settings or with DO_NOT_TRACK=1.
A separate October 15 maintainer issue remains open: the pinned gateway image is amd64-only, preventing model startup on Arm systems including RTX Spark N1X and DGX Spark. The proposed multi-architecture image is a future fix, not a capability delivered by this release.
