0xSero / omarchy-local-ai

What it does

• Validated models per card, or one across several. recipes.json holds, for each supported hardware kind, every recipe accepted on that exact card or across 2 or 4 of them in local-ai-registry: download, load, a correctness check and speed at
• What the machine needs. A recipe that keeps experts in system RAM or reads from disk while it serves (Qwen3.8-Flash-Next on a 3090 or a B70) says how much free RAM and disk it needs and whether the models folder must be on NVMe; it is offer
• Weights are downloaded as you, from the pinned revision, and every file is checked against the hash Hugging Face lists for it before it is used: the SHA-256 of a large (LFS) file, the Git blob SHA-1 of a small one (configs, tokenizers, chat
• Containers. The engine runs on its own internal Docker network that the host has no address on, shared only with its gateway, with no published port and no-new-privileges. A keyed gateway runs as you on the model's own Docker network, speak
• Agents. pi, Claude Code, Codex, OpenCode, omp, Crush, Grok, Copilot and Hermes open in a terminal, in the folder you pick, pointed at the gateway. Nothing in their own config is touched. A model's details list the agents once, the chosen on