How to Use Muse Code: Install Meta’s Coding Agent, Pricing and Muse Wonder (2026 Guide)

Muse Code is Meta’s command-line coding agent, and since 12 October 2026 it is the main way to use Meta’s most capable model, Muse Wonder Full. This guide explains how to install Muse Code on Mac, Linux or Windows, sign in, choose between Muse Wonder and Muse Spark, set safe permissions and pick a subscription, as well as one setting every new user should check: Contributor mode, which lets Meta train on your code and is switched on by default for subscribers.

Commands, prices and settings below were checked on 13 October 2026 against Meta’s developer documentation at dev.meta.ai, Meta’s announcement and independent reporting. Muse Code is updated frequently (the changelog was at version 1.4.6 when we checked), so some commands may change. Prices are in US dollars.

What Is Muse Code?

Muse Code is a terminal-based AI coding agent from Meta. Like Claude Code or Cursor’s agent, you start it in a project folder and describe what you want in plain English; it reads your files, edits code, runs commands and reports back. It launched in beta in August 2026, initially running Meta’s Muse Spark models. It also offers a TypeScript SDK (@muse-code/sdk) for building it into your own tools and can run non-interactively in CI with muse exec.

Muse Code is separate from Meta’s consumer Muse assistant; for that, see our guide to how to use Meta Muse.

Muse Wonder and Muse Wonder Full Explained

Meta Superintelligence Labs released two new models on 12 October 2026:

• Muse Wonder (muse-wonder-1.0, labelled Preview) is available through the Meta Model API and in Muse Code on every plan, including pay-as-you-go.
• Muse Wonder Full has no API and runs only in Muse Code, on the Power Usage and Maximum Usage subscriptions.

According to Meta’s model documentation, Muse Wonder has a context window of about one million tokens and accepts text, images, video, audio and PDFs, with text output. Meta highlights two behaviours: Active Memory, which carries important details across conversations, and Active Focus, which keeps working towards a goal, recovers from errors and checks in before important actions. Meta claims state-of-the-art results on several agentic and coding benchmarks; these are the company’s own claims, and its model card does not publish coding scores.

Muse Code Pricing

On pay-as-you-go, Meta’s API prices muse-wonder-1.0 at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, with no extra charge for long context. Web search grounding costs $2.50 per 1,000 queries. The cheaper muse-spark-1.3 model costs $1.25 input and $4.25 output per million tokens. We found no official free tier for Muse Code.

Meta’s subscription terms say upgrades take effect immediately with prorated billing, downgrades apply from the next cycle, and payments are not refunded unless required by law. You must be 18 or over.

Contributor Mode: Check This Setting First

Meta offers two data modes:

• Standard mode: Meta says your content is not used to train its models.
• Contributor mode: Meta may use your prompts, code, files and responses to train future models, in exchange for lower prices or more usage.

Meta’s documentation says Muse Code subscriptions default to Contributor mode. You can turn it off at any time with the /settings toggle inside Muse Code, after which your future content is not used for training. If you work on client code, proprietary projects or anything covered by a confidentiality agreement, switch Contributor mode off before you start. Contributor pricing is not available for Muse Wonder through the API.

How to Install Muse Code

Step 1: Run the installer

On macOS or Linux, open a terminal and run:

curl -fsSL https://dev.meta.ai/install.sh | sh

On Windows, open PowerShell and run:

irm https://dev.meta.ai/install.ps1 | iex

Then check it worked:

muse --version

Note that Meta’s product page mentions macOS and Windows, while the documentation also gives a Linux installer. On Windows, Meta notes some limits: Muse Code uses PowerShell instead of Bash, the sandbox may trigger a User Account Control prompt, and voice input and session messaging are not available.

Step 2: Sign in

Open your project folder and start Muse Code:

cd your-project
muse

On first run, choose browser sign-in with your Meta Model API account or paste an API key. Type /login to change this later. For scripts and CI, set the key as an environment variable:

export META_API_KEY=your_key_here

An API key takes priority over a browser session. Use muse logout to sign out. A subscription works only through the Muse Code CLI, not through your own API calls.

Step 3: Choose a model

Pick the model for a single run with --model, or switch during a session with /model followed by the model ID (or /models to see the list). For example:

muse --model muse-wonder-1.0

Meta’s documentation still lists muse-spark-1.2 as the default, so check which model is active before a long task. Settings live in ~/.config/muse/settings.json.

Step 4: Turn off Contributor mode if needed

Type /settings and switch off Contributor mode if you do not want your code used for training.

Step 5: Give it a first task

Start with something low risk, such as “explain the structure of this project” or “add tests for the date parsing function”, and review the changes before accepting them.

Permissions and Safety Settings

Muse Code runs commands on your computer, so its permission settings matter. Meta documents:

• Approval modes with --approval-mode: on-request (default), untrusted and never.
• Profiles: Ask me, Auto-review (the default for new sessions), Unrestricted and Read-only.
• Operating system sandboxing: Seatbelt on macOS, bubblewrap on Linux and a Windows sandbox, limiting writes to your workspace and temporary folders.
• Network control with --sandbox-network: proxy-only (default), restricted or unrestricted.
• --yolo disables both approvals and the sandbox. Only use it in a disposable container.

Meta also warns that MCP tools are not sandboxed. Our guides to AI agent security and NVIDIA OpenShell explain how to contain coding agents more broadly.

Memory, MCP, Skills and Subagents

• Memory: Muse Code keeps three memory scopes: a personal project memory, a committed project memory in .agents/memory/ inside your repository, and a personal memory, each indexed by a MEMORY.md file.
• MCP: add servers in the mcp_servers block of your settings or a .mcp.json file, using stdio or streamable HTTP. Use muse mcp login <server> for OAuth and /mcp to list servers.
• Skills and hooks: Muse Code supports skills and hooks and can import skills written for Claude Code and Codex.
• Subagents: up to eight can run concurrently by default.

Muse Code vs Claude Code and Cursor

Muse Code’s low entry price and one-million-token context are its main attractions, while Claude Code and Cursor are more mature and offer IDE integrations. Read our Claude Code guide and Cursor vs Claude Code comparison, and see our guide to X Subscription pricing for how Cursor is now bundled with Grok. For an open-source agent, see DeepSeek Harness.

Availability in the UK and Europe

Meta says availability “may vary by region” and does not list supported countries for Muse Code or Muse Wonder. Its geographic use policy restricts some territories, but it does not name the UK or EU. If sign-up or payment fails in your country, Meta’s regional restrictions are the most likely reason. Because Contributor mode involves processing your code for training, UK and EU businesses should also review Meta’s data terms against their UK GDPR or GDPR obligations before using it on client projects.

Limitations

• Muse Wonder is labelled Preview, and Wonder Full is not available through the API.
• Meta has not published independent coding benchmark scores, so test it on your own codebase.
• There is no official mobile app; Muse Code runs on a desktop computer or server.
• Native IDE extensions are not documented; it is a terminal tool.
• Contributor mode is on by default for subscribers.

Frequently Asked Questions

How much does Muse Code cost?

Subscriptions cost $5, $15, $50 or $200 a month. You can also pay per token with an API key; Muse Wonder costs $4 per million input tokens and $20 per million output tokens.

Is Muse Code free?

We found no official free tier. The cheapest option is the $5 Everyday Usage plan or low-volume pay-as-you-go use.

How do I get Muse Wonder Full?

Subscribe to the Power Usage ($50) or Maximum Usage ($200) plan and use Muse Code. Wonder Full is not available through the API.

Does Meta train on my code?

Only in Contributor mode, which is on by default for Muse Code subscriptions. Turn it off with /settings.

Does Muse Code work on Windows?

Yes, through a PowerShell installer, with some limitations such as no voice input.

Conclusion

Muse Code gives developers a cheap way into Meta’s new Muse Wonder models, from $5 a month or pay-as-you-go, with Muse Wonder Full reserved for the $50 and $200 plans. Install it with one command, sign in, pick your model and, before anything else, decide whether to switch off Contributor mode. Start in a test project with approvals turned on, compare the results with your current coding agent and only then trust it with important code.

Sources: Meta – introducing Muse Wonder; Meta Model API – models; Meta Model API – pricing and rate limits; Muse Code; Muse Code documentation; Muse Code subscriptions; Muse Code authentication; Muse Code permissions; Muse Code – extending; Muse Code changelog; Muse Wonder model card; TechCrunch – Meta releases Muse Wonder.