Add the Local AI panel: the validated local model for your GPU, one button on the bar

[0xSero]

Summary

Local models as one button on the bar, in the V1 scope agreed when #8836 was closed: one validated model per GPU tier, Start/Stop, an agent picker, a keyed Tailscale share, and nothing else. This is the plugin that has been shipping third-party as sero.local-ai (marketplace submission omacom/omarchy-plugin-marketplace#4097, where it passed validation and the security baseline and went through five rounds of security review), brought in as omarchy.local-ai.

• Card: the model validated for the detected GPU, which GPUs were detected and which is in use (a picker when there is more than one), Start with the download size shown first, Stop, Open agent, Share on Tailscale. Every refusal is a sentence on the card.
• Controller (bash, no daemon, 5 verbs): hardware match, a fail-closed recipe gate (digest-pinned image, pinned model revision, no host IPC, no added capabilities, no weakened seccomp, mounts confined to two plugin-owned roots), weights with progress, the engine and an attested gateway as two labeled containers on a private network, and acceptance before "ready": served model id, unkeyed request refused, chat, a decode-speed floor, Messages, Responses, a tool call. Failure rolls back to the previous model.
• Agents are launch-only: pi, omp, opencode, ori, claude, codex, grok, agy, hermes, copilot, crush open on the model with the endpoint and key in their own environment. No user config file is read or written (a test asserts it).
• Share publishes the gateway on the tailnet address, keyed; no tailscale serve, no root, no sudo, no pkexec. The key lives in one 0600 file and enters no process argument.
• Recipes are vendored from the local-ai-registry export: 34 hardware ids, each validated on that card.

How we know it works

• On real hardware, through the plugin's own Start path: 29 of the 31 NVIDIA recipes attempted passed on rented cards (RTX 3060 through RTX 6000 Ada and A6000; the other two had no offers that day), driving the shipped plugin CLI inside the recipe's own image with a docker shim. Per-card JSON with logs and generated command lines: test/rented-results/ in the plugin repo. Total rental cost $0.76.
• On a live Omarchy host (two RTX 3090 + two Arc Pro B70), this exact branch run as the first-party shell from the checkout: the panel added with omarchy bar put omarchy.local-ai, then Stop, Start (accepted at 24 tok/s), Open agent, and Share on/off all driven through the omarchy.local-ai IPC. omarchy local ai … routes through the new bin/omarchy-local-ai.
• Tests: 86 shimmed tests (docker, curl, nvidia-smi, tailscale, agents) run by test/shell.d/local-ai-test.sh; upstream's bin-style, panel-command-path, qml-text-format, plugin-registry-contract, bar-widget-contract, plugins, and manifest-entrypoints tests pass on this branch.

Files

shell/plugins/panels/local-ai/ (Panel.qml, manifest, README, bin, lib, recipes.json, test/all), bin/omarchy-local-ai (router), one line in bin/omarchy for the local group, a section in manual/17-ai.md, a row in shell/plugins/README.md, and test/shell.d/local-ai-test.sh.

🤖 Generated with Claude Code

https://claude.ai/code/session_013f3DyC3B8UpdJKu3Z5F8pX