# How to Use Muse Wonder: The Full Muse Code Setup Guide Meta's new Muse Wonder lives in Muse Code. Here's what Muse Code is and how to get it running with Claude Code doing the setup for you, even if you've never used a terminal. _Updated October 12, 2026._ _Canonical: https://agentnative.inc/resources/how-to-use-muse-wonder_ --- Meta just released a brand new model called Muse Wonder, and some people are saying it's better than GPT-6 Astra and Claude Opus 5.5. The easiest way to use it right now is Muse Code, Meta's coding agent. This is the full guide: what Muse Code is, how to get it set up without touching any code yourself, and how to switch it to Muse Wonder. ## What is Muse Wonder? Two models came out today. **Muse Wonder Full** is the bigger, more expensive one, and **Muse Wonder** is the smaller one. Both are built for agents: tasks with lots of steps that need to remember things and keep going. ```figure src: /resources/muse-code/introducing-muse-wonder.jpg alt: Meta's announcement graphic reading Introducing Muse Wonder, surrounded by rows of the word WONDER. caption: Meta released Muse Wonder and Muse Wonder Full on October 12. source: https://about.fb.com/news/2026/10/introducing-muse-wonder-models/ sourceLabel: Meta ``` So how good is it? It beats Opus 5.5 and GPT-6 Astra at six things: - **Everyday knowledge work and agent tasks.** - **Memory.** The smaller Wonder actually scored the best here. I think this is the model that ends up powering the Muse app. - **Multi-step tasks.** It keeps going when something breaks. - **Very long coding projects.** - **Legal agent work.** It scored the best by far. It's basically the best law agent in the world right now. - **End-to-end business workflows** across apps. ```figure src: /resources/muse-wonder/benchmarks.png alt: Bar charts of six benchmarks. Muse Wonder Full or Muse Wonder is first on the Personal Superintelligence Index, MemoryBench, FocusBench, SWE-Marathon, AutomationBench-AA and Harvey LAB-AA, ahead of Opus 5.5, GPT-6 Astra and Fable 5.1. caption: Scores from Meta's eval table, October 12, at max effort. source: https://research.meta.ai/static/muse-wonder-model-card sourceLabel: Model report ``` It's not the best at everything. Opus 5.5 is still ahead at writing code from scratch and at using a computer. But for long, real-world agent work, Wonder is the one to try. ## What is Muse Code? [Muse Code](https://dev.meta.ai/products/muse-code) is Meta's version of Claude Code. It's an agent that lives in your terminal: you tell it what you want in plain English, and it builds it. It runs on your Mac or PC, works in whatever folder you start it in, and it's where Muse Wonder is available today. You don't need to know how to code. The only two things you'll ever type into the terminal are `muse` and `/model`. ## The setup, step by step Here's the trick that makes this easy: you don't install Muse Code yourself. You ask Claude Code to do it. ```figure src: /resources/muse-code/setup-steps.png alt: Five steps. Open Claude Code with bypass permissions on, ask it to set up Muse Code, type muse and sign in, type /model and pick Muse Wonder, then build. Optionally start with muse --yolo. caption: The whole setup. Claude Code does the installing. ``` ### What you need 1. **Claude Code.** If you don't have it yet, start with [why I use Claude Code instead of Claude](/resources/use-claude-code-not-claude). 2. **A Meta account.** You'll sign in to Meta's developer site, [dev.meta.ai](https://dev.meta.ai), the first time you open Muse Code. 3. **A way to pay.** Either pay as you go (Muse Wonder is $4 per million tokens in and $20 out) or get a Muse Code plan. More on that below. ### Step 1: Open Claude Code with bypass permissions on Bypass permissions lets Claude Code install things without stopping to ask you about every single command. Open your Terminal app and start Claude Code like this: ``` claude --dangerously-skip-permissions ``` ### Step 2: Ask Claude Code to set up Muse Code Copy this and paste it into Claude Code: ```prompt Paste into Claude Code Hey, I want to be able to run Muse Code in my terminal. Please set that up for me. Install it with Meta's official installer from https://dev.meta.ai/docs/muse-code (on Mac or Linux that's "curl -fsSL https://dev.meta.ai/install.sh | sh", on Windows it's "irm https://dev.meta.ai/install.ps1 | iex" in PowerShell). Make sure the "muse" command works in a brand new terminal window, and check it with "muse --version". When you're done, tell me in plain English exactly what to type to start it. I'm not technical, so keep it simple. ``` Claude Code downloads Muse Code, installs it, fixes anything that gets in the way, and checks that it works. When it says it's done, close that terminal window and open a fresh one. ### Step 3: Type muse In the new terminal window, go to the folder you want to build in (or ask Claude Code to make one for you) and type: ``` muse ``` The first time, Muse Code asks two things: 1. **Do you trust this folder?** Say yes. That lets it work on the files in there. 2. **How do you want to sign in?** Pick the browser. It opens [dev.meta.ai](https://dev.meta.ai), you log in with your Meta account, and you're back in the terminal. ### Step 4: Type /model and pick Muse Wonder Muse Code starts on an older Muse Spark model. To switch, type: ``` /model ``` Choose **Muse Wonder** from the list. You'll see it costs more than Spark 1.3, Meta's previous model, but it's a lot smarter on anything with more than a few steps. ```figure src: /resources/muse-code/pick-a-model.png alt: Three models. Muse Spark 1.3 at $1.25 in and $4.25 out per million tokens, Muse Wonder at $4 in and $20 out, and Muse Wonder Full on the $50 a month Power plan or the $200 Maximum plan. caption: Pick Muse Wonder. Wonder Full needs a Power or Maximum plan. source: https://dev.meta.ai/models/muse-wonder sourceLabel: Meta ``` ### Step 5: Turn on yolo mode and build By default, Muse Code stops and asks before it does anything risky. If you want it to just go, the way I use it, quit (Ctrl+C) and start it again with: ``` muse --yolo ``` Yolo mode means no approval prompts and no sandbox, so it can change anything on your computer. Only use it in a folder you don't mind it messing with. Now just tell it what you want, like "build me a landing page for my coffee shop" or "make a simple app that tracks my workouts", and watch it go. ## What does it cost? - **Pay as you go.** Muse Wonder is **$4 per million input tokens and $20 per million output tokens**. Spark 1.3 is $1.25 and $4.25. You get billed for what you use. - **Muse Code plans.** A flat monthly price instead of paying per token. The **Power** plan is $50 a month and is the one that unlocks **Muse Wonder Full**. **Maximum** is $200 a month for 4x more usage than Power. You subscribe in Meta's Accounts Center under Subscriptions, then Muse Code. One thing to know: Muse Code plans start in **Contributor mode**, which lets Meta train on your prompts and answers. If you're working on anything private, check that setting first. ## If something goes wrong - **"command not found: muse".** Close the terminal and open a new one. If it still doesn't work, go back to Claude Code and say "muse isn't found in a new terminal, please fix it." - **Muse Wonder Full isn't in the /model list.** It's only on the Power and Maximum plans. Regular Muse Wonder works for everyone. - **It keeps asking for permission.** That's the default. Use `muse --yolo` if you want it to stop asking. - **You want to sign in again or switch accounts.** Type `/login` inside Muse Code. - **Stuck on anything else?** Paste the error into Claude Code and ask it to fix it. That's the whole point of setting it up this way.