Every plugin here is MIT-licensed and runs entirely on your machine. No account, no cloud, no telemetry. The hosted service is a separate thing you never have to touch.
An Omarchy plugin that cleans up the web. It removes ads, popups and sales offers, says no to tracking-cookie dialogs, blocks trackers, and clears legal popups without agreeing to anything. Filter lists go stale the week they ship, so this reads the page instead: rules and page heuristics act instantly, and a model on your own GPU classifies the rest without sending the page anywhere.
Hold Ctrl+Alt and the page zooms out with every ad outlined in red, so you see what is about to go before it goes. Press Delete and they are gone. Let go instead and nothing happened.
Something got through? Press P and click it. That is a permanent note about that site, in your own hand, and the model is shown it as an example next time.
Turn on automatic mode and ads, popups and offers are removed as each page loads. It is off by default, because deleting things on someone's behalf should be a decision they made.
Consent dialogs are answered, not hidden: it presses Reject, or opens "Your Privacy Choices", switches off everything the site lets you switch off, and confirms. Banners that arrive late are answered too. Every request carries Global Privacy Control, and 68 third-party trackers are blocked before they load.
Terms and privacy prompts that can be declined are declined; ones that offer only "Agree", like a news site's, are removed and the page scrolls again, with nothing agreed to. It never touches a form you are filling in, because hiding an agreement checkbox hides the terms without unbinding you from them.
Ads removed, trackers blocked, dialogs declined and notices cleared, for the page you are on and for all time: in the extension popup, on the bar button, and in omarchy-adblock stats. Counted once when something actually goes, however often a page retries it.
The local backend is the default. When the model runs, it sees a structural sketch of at most 25 elements with a few words each, never the page itself. Private addresses, .local names and your allowlist never reach it at all. omarchy-adblock incognito on lets it run in incognito windows, so private browsing is blocked and remembered too.
The installer asks before adding the extension to Chromium's startup flags, and backs the file up first. It asks again before installing an update hook or linking the command. Nothing runs as root, the default install adds no Python packages, and uninstall.sh takes all of it back.
The plugin adds the bar button; install.sh wires up the blocker in Chromium.
omarchy plugin add https://github.com/OmarchyFans/omarchy-chromium-ad-blocker --enable
cd ~/.config/omarchy/plugins/fans.omarchy.chromium-ad-blocker
./install.sh
bin/omarchy-adblock restartAny OpenAI-compatible server on your machine: llama.cpp, ollama, vLLM, or Omarchy's local agent. Built and measured on a 4B model on a 4 GB laptop GPU. Without one, the rules and heuristics still handle most ads and cookie walls, and Claude Haiku is an opt-in alternative.
Search every Omarchy keybinding, command and manual page without a network. Run what you find, open it, or ask a local GPU model about it in plain words.
A graphing calculator with the face and the feature set of a TI-84 Plus: graphs, tables, lists, regressions, statistics tests and matrices.
Start a coding agent from a keybinding: on this machine, in Docker, or on a GPU machine in the cloud. One dashboard for what is running and what it has cost.
Browse the Omarchy plugin marketplace and audit any plugin in a sandbox before it touches your shell. Plugins run unsandboxed inside a long-lived process; this reads them first.
omarchy plugin add?No, and nothing can be: that command installs Quickshell shell plugins — bar widgets, panels, overlays. An ad blocker is a Chromium extension plus a small native host, so it ships as its own repo with an installer that follows Omarchy's conventions, including a post-update hook so an Omarchy update cannot quietly unwire it.
Whatever your server is serving. It was built and measured against a 4B model on a 4 GB laptop GPU, which is enough: the job is a small classification over a couple of dozen short structural records, not a conversation.
The first two layers are synchronous and free. The model pass happens once per site, in the background, and the answer is written to disk — on every later visit the rules are applied before the page paints.
No. These are MIT-licensed and free, including the one that can point at our paid GPU machines. If you never buy anything they keep working.