Open source · macOS menu-bar agent

Your Mac's local AI, one keystroke away.

Press ⇧⌥Space and say what you need. Muon fixes your writing, explains errors, reads documents and screenshots, and tidies files, using models that run on your Mac. Free and MIT licensed.

Download for macOS Download extension View on GitHub

macOS 26 · Apple Silicon · Apple Intelligence on · or brew install --cask harshdvaid24/tap/muon

Muon's robot mascot — Illustrative · process names and sizes are examples

Muon demo: a spoken request, rewriting an email, explaining an error, answering from a scanned PDF, explaining a pasted screenshot, meeting notes from a recording and totaling receipts — Real recording · Watch in HD

How it works

Muon handles your Mac. The browser agent handles websites.

Muon works on its own. If you also install the browser agent, Muon passes website tasks to it, and it does them in a side panel next to the page.

01 · YOUAskType or say it: “free up ram”, “explain this error”, “fill the signup form on example.com”.

02 · MUONMenu bar, on your MacAnswers repeats from memory, runs Mac tools, and uses Apple's on-device model, or the AI provider you pick (LM Studio by default), when it needs a model.

03a · MAC TASKDone in placeFiles in allowed folders, apps, processes, documents. Anything that moves, deletes or stops asks you first.

03b · WEBSITE TASK · OPTIONALMuon Browser AgentOpens in Chrome or Brave. A planner decides the steps, a navigator clicks and types. Or route it to the browser-use engine.

It only thinks as hard as it has to.

Common requests (writing, documents, screenshots, cleanup, rules) are parsed directly, so they never depend on a model guessing. The rest goes to the smallest tier that can handle it.

Tier 0 · memory~0 ms

Repeated requests replay from a small local cache. No model at all.

Tier 1 · on-device~1 s

Apple's built-in model handles single steps, chat and all writing and explaining. No app memory used.

Tier 2 · local LLMon demand

LM Studio (Qwen 3.5, MLX) by default plans multi-step work, reads images and very long text. Loads only when your Mac can spare it, then unloads. Or pick Ollama, any OpenAI-compatible server, OpenAI, Anthropic or Gemini.

Features

The everyday things you open a chatbot for, without leaving the app you're in.

Muon · Mac app

An email rewritten in a professional tone, with Copy and Paste buttons

Fix your writing

Copy text anywhere, ask “make this professional” or “fix grammar”. Paste puts the result back into the app you came from.

make this professional

Explanation of an error copied to the clipboard

Explain errors and code

Copy a stack trace or a block of code and ask “explain this”. You get what it means and what to do next.

explain this

A pasted lease PDF with a question answered from it

Ask your documents

Drop a PDF, .docx, .csv or image on the palette and ask about it. Scanned PDFs are read with on-device OCR.

who pays the deposit and how much

Explanation of a build error from a pasted screenshot

Screenshots, recordings, receipts

Explain a screenshot, turn a recording into meeting notes, total a folder of receipt photos into a CSV.

explain the error in this screenshot

Confirmation card listing the exact files to move

Files in allowed folders

Move, rename, find duplicates and large files, only inside folders you allow. It lists the exact files and asks once. Delete always means Trash.

move the screenshots in ~/Downloads into ~/Downloads/Archive

Apps listed by memory use, answered from memory in 59 ms

Find what slows your Mac

“what is eating my cpu” lists processes by CPU or RAM. Stopping one shows a confirmation card. System and other users' processes are never stopped.

which apps are using the most memory

Recent requests listed under the suggestions with a highlighted row

Memory of past requests

Anything you've asked before comes back in milliseconds from a capped local cache, with no model at all. It learns your project names too.

A desktop robot

A small draggable robot sits on your desktop. Click it to open Muon, right-click for options. It idles at 8 fps.

The palette listening for a spoken request

Voice, shortcuts, muon://

Press ⌘⇧M and speak. Call it from scripts or Raycast with muon://ask?q=…. Its Mac tools are a standard MCP server.

Muon · more of what it does · real screenshots

Meeting notes generated from a voice recording

Meeting notes from a recording

Summary, decisions and action items from audio or video. Transcribed on-device.

meeting notes from ~/Downloads/standup.mp4

Expense table built from receipt photos with a save CSV confirmation

Receipts into expenses

Merchant, date and total per receipt, the sum, and a CSV saved after one confirmation.

total the receipts in ~/Documents/Receipts

Developer chores

Code search, builds on simulators, git commits, Metro cache resets, crash explanations.

search code for TODO in ~/Projects/weather-app

The in-app catalog of everything Muon can do

Ask what it can do

The whole catalog, inside the app, instantly. No model needed.

what can you do

For you

Old screenshots to archive, how big Downloads has grown, duplicate space to reclaim, a crash to explain. Computed at most once a day, only when you open the palette.

⇧⌥Space

Rule saved confirmation for a weekly cleanup

Rules and undo

A cleanup you have run three times is offered as a weekly rule. Rules run on wake with a notification. “undo” reverses the last automated move.

every monday move the screenshots in ~/Downloads older than 30 days into ~/Downloads/Archive

Muon Browser Agent · browser extension · optional

Planner and navigator

One agent plans the steps, another acts on the page: clicks, typing, scrolling, switching tabs. Progress shows in the side panel.

Record once, replay

Record a workflow by doing it, then replay it later without calling a model.

Any model you like

Use a local model through LM Studio or Ollama, or a cloud provider with your own API key. Pick different models for planner and navigator.

An on-page robot

The same robot floats on the page and shows what the agent is doing. Hide it per site or turn it off.

Privacy and local models

Your text, documents and recordings stay on your Mac.

By default Muon sends nothing to a server unless you ask for something online: a web search, reading a web page, or opening a link. A cloud AI provider is used only if you pick one.

Apple's on-device model firstMost requests are parsed without a model or answered from memory. The rest go to Apple's on-device model.
A bigger local model only when neededMulti-step plans, images and very long text use Qwen 3.5 in LM Studio by default, or Ollama or any OpenAI-compatible server you run. A local model loads only when memory, heat and battery allow, and unloads after five idle minutes.
No idle costAbout 56 MB in the menu bar, no background daemon, no model held in memory.
Cloud is your choiceMuon, the browser agent and the browser-use engine use a cloud model only if you pick one and add your own API key, stored in your Keychain or the extension. With LM Studio or Ollama, everything stays local.

Built so a model can never do something you did not approve.

No shell, ever
Typed tools only, running allowlisted programs with argument lists.
Allowed folders only
~/Library, ~/.ssh, /System and other protected places are always refused.
You approve every change
Moves, renames, quits, builds, commits and saved files ask first. Releases get a system dialog.
Delete means Trash
Nothing is ever permanently deleted. Put Back always works.
The internet is a few named tools
Only web search, page reading and open-in-browser go online, and only when you ask.
Everything is logged
An audit log of every action. Paths only, never file contents.

Models and providers

Local by default. Bring a key if you want the cloud.

Muon (Mac app)

Apple on-device model (local)
LM Studio · Qwen 3.5 (local)
Ollama (local)
Any OpenAI-compatible
OpenAI
Anthropic
Gemini

Apple's on-device model answers most requests. Bigger tasks use what you pick in Settings › AI provider: LM Studio by default (Qwen 3.5 9B, about 6 GB, with 4B as a low-memory fallback, about 3 GB). Outlined in cyan: runs on your Mac. OpenAI, Anthropic and Gemini need your own key, kept in your Keychain.

Browser agent (extension)

LM Studio (local)
Ollama (local)
OpenAI
Anthropic
Gemini
DeepSeek
Grok
Azure OpenAI
OpenRouter
Any OpenAI-compatible

Outlined in cyan: runs on your machine. Others need your own API key.

Compared

Built to act on your Mac without handing a model the keys.

The local agents people run on a Mac, from each project's own README on 28 September 2026. Muon is the newest and smallest of them; this is what it is built to be.

	Muon	OpenClaw	goose	Open Interpreter	AnythingLLM · Jan
What it is	Menu bar agent for everyday Mac work	Personal assistant in your messaging apps, persistent Gateway daemon	General-purpose agent, desktop + CLI, MCP extensions	Coding agent forked from Codex	Private chat and RAG apps
How the model changes things	Never a shell. Typed tools, allowlisted binaries, path policy, one card per change, Trash not delete	“Tools run on the host … unless you configure sandboxing”	Shell and files via its developer extension; approval configurable	Commands in native sandboxing, with approvals	Do not act on the Mac
Everyday ask	1–3 s on Apple's on-device model; repeats in ms	Model turns, your provider	Model turns, your provider	Model turns, your provider	Local chat, seconds or more
Idle cost	56 MB, no model, no daemon	Gateway daemon always on	Desktop app	CLI while running	Model loaded while open
Needs a big model	No	Hosted or local	15+ providers incl. Ollama	OpenAI-compatible providers	Yes, local
Your files and recordings	Stay on the Mac	Depends on provider	Depends on provider	Depends on provider	Stay on the Mac
Install	Download or brew cask	curl \| bash or npm	Desktop app or script	curl \| sh	Download

Not for: hour-long autonomous tasks, chatting over WhatsApp, or being your coding agent. For that, Muon opens Claude Code, Codex, Gemini CLI or Aider in a VS Code window for the project you name.

Install

Start with the Mac app. Add the rest only if you want it.

Needs macOS 26 on Apple Silicon with Apple Intelligence on, and Node.js 20+ (brew install node). Optional: a local model for multi-step planning and images (step 2), Laya for faster routing. Check your setup with Muon --diagnose.

Step 1 · required

Muon for Mac

Download Muon.zip and move Muon.app to Applications, or use Homebrew. Needs macOS 26 on Apple Silicon with Apple Intelligence on, and Node.js 20+.

brew install --cask harshdvaid24/tap/muon

Muon isn't notarized yet. On first launch, open System Settings › Privacy & Security and click Open Anyway.

Download for macOS

Step 2 · optional · local model

Download the model with LM Studio

For multi-step plans, images and very long text. Install LM Studio, then get Qwen 3.5: 9B (about 6 GB) and the 4B low-memory fallback (about 3 GB).

Get LM Studio

lms get qwen/qwen3.5-9b@4bit --mlx -y
lms get qwen/qwen3.5-4b@4bit --mlx -y

Or search for them in LM Studio: qwen3.5-9b, qwen3.5-4b.

Muon › Settings › AI provider can also download the model for you.

Prefer Ollama? Get Ollama, run ollama pull qwen3:8b, then pick Ollama in Settings › AI provider.

Step 3 · optional

Browser agent extension

For website tasks. Works in Chrome, Brave and Edge.

Download the zip from the latest release and unzip it to a folder you will keep.
Open chrome://extensions and turn on Developer mode.
Click Load unpacked and pick that folder.

Open the side panel with Alt+Shift+B, then choose a model in Settings.

Download extension Build from source

Step 4 · optional

browser-use engine

A single-loop agent that runs in its own browser profile. Muon uses it for website tasks when it's installed, and falls back to the extension if it fails.

uv sync
./install-native-host.sh

Then pick it in Muon › Settings › Browser tasks, or in the extension under Settings › General › Agent engine.

Or · from source

Build it yourself

Xcode 26 and Node 20+.

git clone https://github.com/harshdvaid24/muon.git
cd muon/mac-tools && npm install && npm run build
cd .. && scripts/build.sh && scripts/run.sh

Ask it anything

Plain language. No syntax to learn.

A few things people ask. The full catalog covers writing, documents, screenshots, voice, expenses, crashes, developer chores, files, disk, apps and automation.

fix grammar
shorten
reply saying I'll be there at 5
extract the action items
summarize ~/Documents/lease.pdf
read the latest screenshot
rename my screenshots
transcribe ~/Downloads/call.m4a
why did my app crash
commit this
run weather-app on iPhone 17
clean the metro caches
find duplicate files in ~/Downloads
which apps are using the most memory
in safari open a new private window
undo

Open source

Read the code. Fork it. Built on good work by others.

MIT