Omarchy will run local AI models for your coding agents from a curated recipe catalog
One command picks an open model that fits your hardware, runs it in a locked-down Podman container with no network access, and puts it in the agents panel next to your subscriptions.
Image: omacom/omarchy pull request #17267 (GitHub)
Your coding agents can soon run on a model that never leaves your machine. DHH merged pull request #17267, “Run local models for agents from a curated recipe catalog”, into Omarchy’s quattro branch on November 1, 2026. The headline command is omarchy agent local use best, which, according to the PR, “runs an open model on this machine for your coding agents, picked for the exact hardware from a curated recipe catalog”. It is coming in a future version of Omarchy.
The recipes are where the curation happens. Each one pins its runtime image by digest and its Hugging Face weights by revision, verified by hash, and records measured memory use for each context size along with benchmark numbers per hardware class. A best.toml file maps GPU family and RAM class to the right recipe, and setup takes the largest measured context that fits the machine.
Security got serious attention. Every model runs as a rootless Podman container on an internal network with no egress, a read-only root filesystem and all capabilities dropped, with only allowlisted devices. The API key is stored as a Podman secret, and after start-up Omarchy checks that the server refuses unauthenticated requests and serves the model the recipe promised.
You can run several models at once, and each gets its own line in the agents panel next to your subscriptions: the maker’s logo, play/pause, a terminal button and a memory meter, plus progress bars for loading, stopping and removing. Memory checks count the models already running, and the play button is greyed out when a model won’t fit. “Add another” now offers Local AI with the recipes that fit your machine, each card showing architecture, memory, context and measured speed, for example “Dense · 58 GiB Memory · 256K Context · 14 TPS”.
To put a model to work, omarchy agent --local or --local-model <recipe> launches the default agent, OpenCode or pi, on it. Both runtimes serve Chat Completions, Responses and Anthropic Messages, so other agents can connect too. Separately, the new omarchy agent --agent <id> powers a terminal button on each subscription, opening Codex for OpenAI and Claude Code for Anthropic.
DHH measured the first two recipes on a Ryzen AI Max+ 395 with 128 GB of memory, both at their trained 256K context. Qwen3.8 Flash Next, a MoE model, peaks at about 99 GiB and decodes at 41.3 tokens per second on a 32K prompt. The dense Qwen3.8 27B needs about 55 GiB and decodes at 14 tokens per second on the same prompt.
The engine and the recipe catalog ship together as the omarchy-local-ai package, built from omacom/omarchy-local-ai. It was added to Omarchy’s package repository by omarchy-pkgs #1979 and moved to version 0.2.1 by omarchy-pkgs #1982, which made Podman and crun optional. The package becomes part of Omarchy’s base install so the panel only offers Local AI on machines a recipe fits, and omarchy agent local use installs Podman and crun the first time you set up a model. Catalog updates then arrive with omarchy update.
A few things were still awaiting a manual check when the PR merged, according to its test plan: the drag preview, the Removing bar on a model that has to stop first, and “Add another › Local AI” on a machine without the package. The end-to-end run on a 128 GB Strix Halo, covering setup, switching, rollback and removal, is ticked off.
