Meta’s Muse Code for UK developers: cost, data and safe set-up

Muse Code is Meta’s answer to Claude Code and Codex: a coding agent that lives in your terminal, runs on Meta’s Muse Spark models and costs from $5 a month. The price is the headline. The setting that matters more is which model it starts on, because when we installed it the default was the tier Meta trains on.

The answer in brief

• What it is: a terminal and CI coding agent from Meta that plans, edits and runs commands, with approvals and an operating-system sandbox on by default.
• Status: out of beta since 31 August 2026, on macOS, Linux and Windows.
• Cost: plans at $5, $15 and $50 a month on Meta’s product page, or pay-as-you-go through the Meta Model API (checked 4 October 2026).
• UK access: yes. The UK is not a restricted territory, and we ran it from the UK on 27 September and 4 October 2026.
• Watch out for: with no model saved, versions 1.4.0 and 1.4.2 both started on muse-spark-1.3-contributor, the model Meta can train on, and 1.4.2 started at max reasoning effort. Meta’s own API terms forbid sending it confidential code or personal data, so switch to muse-spark-1.3 first.
• Our view: worth a trial on personal projects. Don’t use it on client repositories until you have fixed the model, the permission profile and the data question.

What is Muse Code?

Meta describes Muse Code as its “coding agent for the terminal and CI, built for Muse Spark”. You run muse in a project folder and give it a task. It plans the work, edits files and runs commands, stopping for your approval on anything risky.

It runs two ways from the same binary. muse opens an interactive terminal session with slash commands. muse exec "<prompt>" runs one task to completion for scripts and CI pipelines. There is also a TypeScript SDK, in developer preview, for driving sessions from your own program.

Meta’s pitch rests on three things:

• Many agents at once. A lead agent can hand bounded tasks to subagents. They share your checkout unless the lead asks for an isolated Git worktree for a child, which Meta recommends whenever parallel children write files. By default one agent tree runs eight agents at a time, including the lead; you can set 1 to 64, and an unconfigured session at the ultra effort level uses 64.
• An audit trail. Every session is an append-only event log, so you can see which agent did what, and resume or replay a run.
• A model trained with the tool. Meta says Muse Spark was “co-trained” with Muse Code, so tool calls succeed more often than in a generic wrapper. That is Meta’s claim; we have not benchmarked it.

When did Meta release Muse Code?

• 5 August 2026: Meta released Muse Spark 1.2 and “an early beta of Muse Code” in its launch post.
• 31 August 2026: Muse Code came out of beta, adding messages between sessions, multi-agent workflows, rewind, the SDK preview and subscription plans.
• 2 September 2026: Meta released Muse Spark 1.3, which Muse Code can use.

The installer gave us version 1.4.0 on Windows on 27 September 2026 and version 1.4.2 on Ubuntu under WSL on 4 October 2026. Meta’s public changelog lists 1.4.0, 1.4.1 and 1.4.2 as of 4 October 2026, but gives no release dates. Check the version you actually have with muse --version.

How much does Muse Code cost?

You can pay a flat monthly fee or pay per token. Meta’s public prices are in US dollars, the changelog says /usage and /models label costs as USD, and dollars were all the dashboard showed us when signed in from the UK on 4 October 2026. The plan prices below were checked on 4 October 2026.

Source: Muse Code product page and subscription docs, checked 4 October 2026.

Five details from the subscription docs are worth knowing before you sign up:

• Region. “Subscription benefits and availability may vary by region.” Meta shows what is available during onboarding. Signed in from the UK on 4 October 2026, we saw prices in US dollars only, so the cost in pounds moves with the exchange rate.
• One key only. A plan covers the Muse Code API key created during sign-up. Any other key on the same Meta Model API account is billed pay-as-you-go.
• Limits. If you hit your limit you can upgrade or wait. Power Usage users can only wait. Besides the 5-hour window, the 1.4.0 changelog mentions a weekly usage window; Meta does not publish its size.
• Refunds. There are “no refunds for canceled subscriptions” unless the law requires one.
• Data. The plan does not decide what Meta can train on. The docs say that “depends on the models you select”.

Pay-as-you-go runs through the Meta Model API with no subscription. Meta’s pricing page lists two tiers:

Prices per million tokens. Source: Meta Model API pricing, checked 4 October 2026. The 1.2 models are priced the same.

Two things push a pay-as-you-go bill up. Muse Spark’s reasoning tokens are billed as output, and Muse Code may start at the most expensive setting. Meta’s configuration docs give high as the default reasoning effort, but with nothing saved in its settings, our 1.4.2 install started at max, the deepest level, shown in gold on the status line. A 1.4.0 changelog entry points the same way: first-login sessions on 1.3 models “now show max reasoning effort”. Muse Code also runs four background “observer” agents by default (memory recall, skill recall, goal tracking and verification), and Meta’s docs say each one “makes its own model calls”. Use /effort to set a lower level; Muse Code saves it for later sessions.

Which model does Muse Code start on?

Not the one the docs say. Meta’s configuration docs say “The default model is muse-spark-1.2”. In our tests, Muse Code starts on muse-spark-1.3-contributor whenever no model is saved, and says so on screen.

• Windows, version 1.4.0, 27 September 2026: it started on Contributor, including after we removed the model setting from its settings file.
• Ubuntu under WSL, version 1.4.2, 4 October 2026: on first sign-in it reported “Model set to muse-spark-1.3-contributor” and wrote that model into a new settings file. With the settings cleared again, it started on Contributor at max reasoning effort and saved neither choice.

That fits what Meta said at launch: “Muse Code starts on a contributor tier”. The Contributor model is cheap because you let Meta train on your prompts and completions. The on-screen notice is broader than you might expect: “Your content, including inter-session messages, may be used for product improvement.”

The good news is that the fix sticks. Switch before you open anything you would not publish:

1. Type /models in a session and pick muse-spark-1.3, or start with muse --model muse-spark-1.3. In version 1.4.2 the picker showed the training notice beside each Contributor model.
2. Muse Code saves the choice to ~/.config/muse/settings.json. When we signed out and back in on 4 October 2026, it kept muse-spark-1.3. If you edit the file by hand, keep "schema_version": 1 in it or every command fails at startup.
3. Check the status line at the start of each session, and again after an update.

Can UK developers use Muse Code?

Yes. Meta’s Geographic Use Policy (last updated 18 September 2026) lists the Restricted Territories where the Meta Model API is unavailable. Neither the UK nor any EU country is on it, and the only model-specific restriction is Muse Image in India. We installed and ran Muse Code from the UK on 27 September 2026 (Windows) and 4 October 2026 (Ubuntu under WSL).

Meta’s policy also says it “may enable or restrict access in any jurisdiction at any time”, so treat this as a snapshot. Our Muse Spark 1.3 guide covers UK access for the model in more detail.

What does Meta do with your code?

It depends on the model, not the plan. Meta’s Data Commitments say:

• Standard models (muse-spark-1.3, muse-spark-1.2): Meta “does not use your Content from Standard Services to train Meta Models”. It can still process content to detect abuse and protect safety.
• Contributor models (muse-spark-1.3-contributor, muse-spark-1.2-contributor): Meta may use your inputs and outputs to “train, develop, evaluate, and improve” its models, products and services.

The Meta Model API terms (last updated 2 October 2026), which also cover Meta’s “coding harness”, turn that into rules:

• Contributor is off-limits for confidential work. Section 6.2 says you “must not submit sensitive, confidential, or personal information” to the discounted services, and that code you are required by contract to keep confidential must not go there either. For agencies and freelancers, that rules out client repositories on Contributor.
• Standard comes with processor terms. Under section 5.1, Meta processes Standard content as your processor under its Global Processor Terms and authorises the subprocessors on a published list. Customers in the European Region contract with Meta Platforms Ireland.

Standard is the necessary choice for client work, not the whole answer. The terms we read do not say which country processes a given Muse Code request, so check the processor terms and subprocessor list against your own UK GDPR obligations before you send personal data through the agent. Meta also offers Zero Data Retention to qualifying accounts through its sales team.

How safe is Muse Code on your machine?

The defaults are careful, with a few settings to tighten. Meta’s permissions docs describe two independent layers:

• Approval. Before a tool call with side effects, Muse Code checks it against a policy. It splits compound shell commands into stages and holds the first risky one: in wc -l report.log && echo cleaning && rm -rf report.log, only the rm -rf stops for review. If you reject it, nothing runs, not even the safe stages before it.
• Sandbox. Shell commands run inside an operating-system sandbox (Seatbelt on macOS, bubblewrap on Linux, a Windows sandbox on Windows). The agent can write to the workspace and a temp folder only, and its own .git, .muse and .agents folders stay read-only. Network access is approved per destination by default.

We tested this on 4 October 2026 with the same pattern in a throwaway project. Muse Code held the command at “Stage 3/3” on rm -rf build, showed the exact arguments, and offered to allow it once, allow rm in this workspace, or abort. We aborted, and it reported “no changes made”.

Four defaults deserve a second look:

• Auto-review is the starting profile. An automated reviewer, not you, decides eligible approval requests. Choose “Ask me” with /permissions to decide every eligible request yourself (routine safe actions still pass without a prompt). Muse Code saves the choice as the default for new sessions and muse exec runs, but a headless run has no one to answer a prompt, so a held action there stays unresolved.
• Hooks and MCP servers sit outside the sandbox. Hook commands run “directly through your shell”, and MCP tools “are not sandboxed”. Only add ones whose code you have read.
• Committed memory loads even in untrusted folders. Meta tells you to treat a repo’s MEMORY.md as “a prompt-injection surface”.
• --yolo removes both layers and trusts the workspace. Meta says to use it only in a disposable, isolated environment, “never” on a workstation with real credentials.

Muse Code is also conservative with your repository. The docs say it won’t commit, amend or push unless you ask in the session.

Will Muse Code pick up your Claude Code set-up?

Yes, and that is worth knowing before you run it in a project you already use with Claude Code. Muse Code reads CLAUDE.md as project instructions: at each folder level it uses the first of AGENTS.md, CLAUDE.md, .agents/AGENTS.md and .claude/CLAUDE.md that exists. It also discovers personal skills in ~/.claude/skills and Codex skills by default, and it scans a repository’s .claude/skills folder. When we started it on Windows on 5 October 2026, it also announced it was “Including your Codex personal rules” from our existing Codex set-up, with /settings to manage them.

For switchers there is a built-in /migrate skill that imports Claude Code or Codex memory notes and MCP servers, and muse skills import --from claude brings in skills. Project instructions, skills and hooks load only after you trust the workspace.

Does Muse Code work on Windows?

Yes, with gaps. Meta’s overview lists four differences on Windows:

• the agent runs commands through PowerShell rather than Bash;
• the sandbox may show a User Account Control prompt the first time it starts;
• voice input is unavailable, although the 1.4.0 changelog says it is now on by default on Windows x64 and ARM64, and it worked in our test (see below);
• session messaging is unavailable.

We tested voice input on Windows 11 on 5 October 2026 with version 1.4.2 in PowerShell. It needed no set-up: Alt+V started recording from the microphone, a level meter appeared in the status line, and Muse Code typed what we said into the prompt without sending it. Its /voice status check reads “audio is sent to internal ASR; mic opens only while recording”, so your speech leaves the machine whichever model you use, and once you send it, the text is treated like any other prompt on the tier you chose. We checked: /voice status reports the same internal ASR route on both muse-spark-1.3 and the Contributor model. Meta’s docs do not say how the audio itself is handled.

Running Muse Code in Ubuntu under WSL gets you the Linux build instead, with Bash and the bubblewrap sandbox. We installed version 1.4.2 that way on 4 October 2026 and tested sign-in, model selection and approvals; we did not test voice input or session messaging there. One difference to expect: in our test, “Log in with browser” reported “Couldn’t open a browser” and fell back to a device code. You open the auth.meta.com link it prints, check the code matches and approve.

Mac users should note that the current Apple silicon package does not include the workflow engine, so multi-agent workflows are unavailable there. Meta’s product page says “Available for MacOS and Windows”, while its docs also cover Linux.

How do you install and remove Muse Code?

Install with Meta’s one-line installer:

• macOS and Linux: curl -fsSL https://dev.meta.ai/install.sh | bash
• Windows (PowerShell): irm https://dev.meta.ai/install.ps1 | iex

Meta’s overview pipes the script to sh, but it is a Bash script, and Meta’s own blog and product page pipe it to bash, so use bash. Then run muse in a project folder. On first run it asks whether to trust the workspace and how to sign in: in the browser or with an API key. For CI, set META_API_KEY. Meta Managed Account users must use an API key.

On Linux and WSL, the installer puts a small muse launcher and the program (about 330 MB for version 1.4.2) in ~/.local/bin, and adds that folder to your PATH in .bashrc if it is missing. Read the script first if you prefer: it is plain Bash and checks the launcher’s SHA-256 when Meta’s server sends one.

We found no uninstall command in Meta’s docs, changelog or installer scripts. muse logout removes the stored sign-in, but not a META_API_KEY you exported yourself. On Windows, the installer we ran on 27 September 2026 put a 424 MB program in %LOCALAPPDATA%\Programs\muse and added it to your user PATH. Settings and sign-in went to ~\.config\muse, and sessions, skills and plugins to ~\.local\share\muse. It also left sandbox folders under AppData\LocalLow and muse-workspace-probe folders under %TEMP%. Removing those folders and the PATH entry, which is what we did, also deletes your local sessions. Revoking the API key on the Meta Model API dashboard and cancelling any subscription are separate steps.

AIWIZ verdict

Muse Code is a serious tool, not a demo. The sandbox, staged approvals and event log are thoughtful, and a $5 plan makes it cheap to try.

It is not ready to point at client work out of the box. In two installs a week apart, it started on the Contributor model, which Meta’s own terms say must not receive confidential code. Its configuration docs still name a different default, and Meta does not say which country processes a given request. Try it on your own projects on the Everyday plan or Standard pay-as-you-go. Before any client repository, switch to muse-spark-1.3, choose “Ask me” if you want to decide approvals yourself, and check Meta’s processor terms against your own data obligations.

For the model underneath, read our Muse Spark 1.3 guide. For the rest of Meta’s Muse family, see What is Meta Muse?.