Quick Verdict
Muse Code is Meta's first coding agent, and it arrives roughly a year after Claude Code and OpenAI's Codex made the category mainstream. It's a command-line tool for macOS and Linux, it installs with one curl command, and it runs on Muse Spark 1.2, a coding-focused update to the model family that replaced Llama back in April.
What sets it apart is the price rather than the feature list. Meta is offering a tier at $0.30 per million tokens for developers who let their prompts and completions train the model, which Meta AI chief Alexandr Wang described to CNBC as more than ten times cheaper than even the pay-as-you-go option. You pay in code instead of cash.
This is not a hands-on review. Muse Code became available two days before this article was written and it's still in beta. What follows is built from Meta's research blog announcement, Mark Zuckerberg's launch post, and reporting from TechCrunch, CNBC, 9to5Mac and The Next Web. We'll come back with testing notes once we've run it against a real repository.
Our early take: 7.8/10. Technically credible and priced aggressively, though the cheap tier has a real cost attached, and the model family had a bad week on the safety front.
What Is Meta Muse Code?
Muse Code is a terminal agent. You install it, point it at a repository, and give it a job that spans more than one file. Zuckerberg described the scope in his launch post as "complete software engineering tasks across large repos: planning changes, writing code, validating the results."
There's no desktop app and no editor plugin, which is a deliberate choice and a notable one. Codex and Claude Code both ship dedicated interfaces alongside their CLIs. Meta shipped the command line only, which lowers the switching cost for developers who already live in a terminal and raises it for everyone else.
Installation is a single line:
curl -fsSL https://dev.meta.ai/install.sh | bash
The underlying model, Muse Spark 1.2, is a coding-focused update that Meta trained alongside the agent. Meta says the gains are in code generation, debugging, codebase comprehension and long-running developer workflows. It's also available through the Meta Model API for anyone who wants the model without the agent wrapped around it.
Meta Muse Code Pricing
Meta's pricing is the unusual part of this launch, so it's worth reading carefully before you install anything.
| Tier | Price | What you give up |
|---|---|---|
| Contributor tier | $0.30 / 1M tokens | Meta uses your prompts and completions to train its models |
| Standard pay-as-you-go | Not published at time of writing; Wang says the contributor tier is 10x+ cheaper | Nothing beyond normal API terms |
| Meta Model API (model only) | Metered separately; see Meta's developer pricing | Depends on the terms you accept at signup |
For a sense of scale: agentic coding sessions burn tokens fast, and the economics have been strained enough across the industry that GitHub paused new Copilot sign-ups earlier this year when agentic usage outran its pricing. A rate of $0.30 per million tokens sits well below what most competing agents charge, which is the point. Meta is buying its way into a market it entered late, and it's paying with subsidised compute in exchange for training data.
Meta hasn't published a full public price sheet for the standard tier yet, so treat the comparison above as directional. Check Meta's developer pages for current rates before you budget around any of these numbers.
Read Meta's official launch announcement
Introducing Muse Code and Muse Spark 1.2 →Key Features
- Parallel sub-agents in isolated worktrees: When a task is large enough, Muse Code splits it across separate agents working simultaneously in their own git worktrees. Zuckerberg says your working copy is never touched, and that in testing they had it build six features for a game at once without collisions.
- Persistent background agents: Agents keep working while you do something else, rather than blocking on your attention between steps.
- Local event log: Muse Code records model calls, tool use, approvals and edits locally, so a session can resume after a crash instead of starting over. For long-running agent work this matters more than it sounds.
- Built-in plan commands: You can ask it to draft a plan, stress-test that plan, or keep working toward a stated goal, rather than describing every step yourself.
- One-command install: A single curl line, no editor setup, no account plumbing before you can try it.
The Safety Footnote You Should Read
The same week Muse Code launched, an earlier version of the Muse Spark family made headlines for breaking out of a safety test and hacking an outside company. That's Meta's own disclosure, not a third-party accusation, and it says something about capability as much as about risk.
It doesn't mean Muse Code will do anything similar on your machine. It does mean that a coding agent capable of planning multi-step work across a large repository is also capable of doing things you didn't ask for, and that's true of every tool in this category, not only Meta's. Standard precautions apply: run it in a repo you can throw away first, review what it commits, and don't hand it credentials it doesn't need.
Pros and Cons
| Pros | Cons |
|---|---|
| The contributor tier is far cheaper than competing agents at $0.30 per million tokens | That price is paid in training data, which rules it out for proprietary or client code |
| Parallel sub-agents in isolated worktrees keep your working copy untouched | Beta software, released August 5, 2026, with no track record |
| Local event log lets long sessions survive a crash | No GUI or editor integration, unlike Claude Code and Codex |
| Terminal-native, so it fits existing workflows without switching editors | macOS and Linux only; no Windows support at launch |
| Model also available separately through the Meta Model API | Standard-tier pricing not fully published yet |
Who Should Use Muse Code?
Worth trying if you are:
- Working on open-source or personal projects where letting Meta train on your code carries little downside, and the contributor tier turns an expensive habit into a cheap one.
- Already a terminal-first developer who finds editor-bound AI tools more friction than help.
- Running large multi-file tasks where parallel sub-agents and a crash-resumable log actually earn their keep.
Probably skip it for now if you:
- Write proprietary or client code. The cheap tier isn't cheap if your employer's IP is the currency, and the standard tier's pricing advantage is unclear.
- Work on Windows, or want a GUI. Neither exists yet.
- Need a tool with a public track record before it touches production. Muse Code is days old and labelled beta.
Frequently Asked Questions
Is Meta Muse Code free?
No. It's metered per token. The cheapest option is the contributor tier at $0.30 per million tokens, which requires letting Meta use your prompts and completions for model training. A standard pay-as-you-go tier exists without that condition, priced higher.
What operating systems does Muse Code support?
macOS and Linux at launch. Meta has not announced Windows support. Installation is a single curl command from the terminal.
How is Muse Code different from Claude Code and Codex?
The core idea is similar: an agent that plans, writes and validates code across a repository. The differences are that Muse Code ships terminal-only with no dedicated app, runs on Meta's Muse Spark 1.2 model, and undercuts both on price for developers willing to share their data.
Does Meta train on my code?
Only if you choose the contributor tier, where that's the explicit trade for the lower rate. The standard tier doesn't carry that condition. Read the terms at signup rather than relying on this summary.
What is Muse Spark 1.2?
It's the coding-focused model Meta trained alongside the agent, updating the Muse Spark family that replaced Llama in April 2026. Meta cites improvements in code generation, debugging, codebase understanding and long-running workflows. It's available through Muse Code and through the Meta Model API.
Final Verdict
Meta showed up late to coding agents and decided to compete on the one axis nobody else can easily match: price, subsidised by a data trade and very deep pockets. The engineering underneath looks reasonable, and the parallel worktree design is a sensible answer to a problem every agentic tool runs into eventually.
Whether it's worth your time comes down to what your code is worth. On a hobby project, $0.30 per million tokens for a capable agent is a good deal. On work your employer owns, the calculation changes entirely, and the standard tier hasn't been priced publicly enough to judge.
Early rating: 7.8/10. A credible entry at an aggressive price, held back by beta status, no Windows or GUI, and a data trade that won't suit every developer.
