Muse Code: Pricing, Install and an Honest Review of Meta's Coding Agent

Muse Code is Meta's terminal coding agent, built around its Muse Spark models. You install it with one command, and plans cost $5, $15 or $50 a month. It runs parallel subagents in separate git worktrees and logs every action for replay. One catch: it starts on the contributor model, where Meta may train on your code.

Meta launched Muse Code in beta on August 5, 2026 alongside Muse Spark 1.2 (Meta), and took it out of beta on August 31 with subscription plans, workflows and an SDK (Meta). Since then it has picked up Spark 1.3 and a surprisingly large set of community tools. This page covers what it costs, how to set it up safely, and what people have actually built with it.

How much does Muse Code cost?

Three monthly plans, from Meta's Muse Code product page:

Plan	Price	Usage	Extras
Everyday	$5/month	"10-50 prompts every 5 hours, including image and video uploads"	Latest Muse models, voice mode, web search
High	$15/month	5x Everyday	More multimodal input
Power	$50/month	20x Everyday	Early access to features, multi-agent coordination by default

Usage is metered by tokens in a rolling 5-hour window, not by request count (Meta). Muse Code isn't included in a Muse agent subscription; see Muse pricing for the agent plans.

How far does $15 go? Federico Ramirez, who uses it for GitHub issue-to-PR work, found the High plan good for "around 2 to 4 hours" of coding before the 5-hour limit, and about 25 hours a week. That's one developer's workload, but it's the only hands-on number we've seen.

You can also skip the subscription and sign in with a Meta Model API key, paying per token: $1.25 per million input and $4.25 per million output on Standard, or $0.10 and $0.20 on the contributor tier (Meta pricing). Several older reviews, including MakerStack's and Layer3 Labs', were written during the beta and describe only per-token billing. The plans came later.

How do you install Muse Code?

From Meta's Muse Code docs:

1. Install. On macOS or Linux run curl -fsSL https://dev.meta.ai/install.sh | sh. On Windows PowerShell run irm https://dev.meta.ai/install.ps1 | iex.
2. Check it. Run muse --version.
3. Start in your project. Run muse inside the repo and accept the workspace trust prompt. Trusting a workspace loads its skills, rules and hooks, so only trust repos you know.
4. Sign in. Pick browser sign-in or an API key. For CI or other non-interactive use, set META_API_KEY. You can reopen sign-in later with /login.
5. Switch the model before you open private code (next section).

DataCamp notes the installer drops a 230 MB binary at ~/.local/bin/muse. The sandbox sets itself up on first run, and on Windows that can trigger UAC prompts.

How do you switch Muse Code off the contributor tier?

Run /model inside a session and pick muse-spark-1.3 (the Standard model) instead of muse-spark-1.3-contributor.

This matters. When DataCamp signed in for the first time, the session header read "Model set to muse-spark-1.3-contributor" with a note that "your content, including inter-session messages, may be used for product improvement" (DataCamp). Meta's own launch post confirms the contributor model "launches as the default." After switching, the "Discounted tokens" notice disappears from the header.

Meta's docs page currently lists muse-spark-1.2 as the default, which doesn't match what DataCamp saw. Check your header rather than trusting either. The New Stack summed up the trade in a headline: "Meta's new coding agent is cheap (but it'll cost you your data)" (The New Stack). Don't run the contributor model on client code, anything under NDA, or repos with secrets. Agencies weighing this should read Wavect's buyer guide.

What can Muse Code do?

• Subagents in their own worktrees. Muse Code splits work across child agents, and "each child gets its own git worktree, so parallel children never collide on the same files" (Meta).
• Replayable event logs. "Every action lands in a replayable per-session event log you can audit." Someone has already built Muse Monitor, which reads those logs and redraws a multi-agent session as a live tree.
• Bundled skills. /plan, /grilling, /grill-with-docs and /taste. They're explicit-invocation only and load for the turn you call them.
• Workflows. Say "use a workflow" and it runs subagents in parallel; watch them in /workflows, and save a finished workflow to reuse (Meta).
• Rewind. Double-tap Esc to roll the conversation back to a safe point from the event log.
• Inter-session messaging over local Unix sockets, so two terminals can pass warnings and answers.
• An SDK (TypeScript, developer preview) that talks the Muse Session Protocol locally.

That SDK and muse serve explain the flood of front ends: Muse Desktop for macOS and Windows, muse-acp for Zed and JetBrains, and several VS Code extensions. See our coding and dev tools shelf for the rest.

Which models does Muse Code use?

Muse Spark 1.3 and 1.2, each in Standard and contributor versions (Meta). Spark 1.3 has a 1,048,576-token context and a max reasoning mode, and Meta claims it uses ~20% fewer tool calls and ~25% fewer tokens than 1.2 (Meta). Meta says Spark 1.2 was "co-trained with the harness," so Spark is at its best inside Muse Code.

The harness isn't locked to Meta models. Ollama showed it running local and cloud Ollama models with ollama launch muse, which is one way to pair it with Muse Glimmer.

What have people built with Muse Code?

Costs below are as reported by each builder.

• A site port for 20 cents. DHH had Muse Code convert the TanStack version of a new site to Astro. He says it did a flawless job.
• Hand tracking in under an hour. Dilmer Valecillos used Spark 1.3 in Muse Code to add hand tracking to his VR mini golf game in under 60 minutes, including a matching toon shader.
• A game remake for $0.63. Moss AI Studio remade Spooky Clicker and published the changelog.
• Five apps for about $10. Ramanpal Singh built a company site, a puzzle game, a 3D traffic sim and more. His catch: Spark 1.2 often failed to enforce game rules like win conditions.
• Research tooling. Kilian Lieret built a benchmark trajectory viewer with it, and Max Weinbach built a Bloomberg terminal-style work page.

One more worth knowing. When Muse Code wouldn't sign in from Docker, the Cline team ported its system prompt into Cline and reports Spark 1.2 fixed a real bug with 2.7x fewer tokens and at less than half the cost ($7.69 to $3.25). The system prompt carries a lot of the weight.

Muse Code review: what independent testers found

The praise is consistent: it's cheap, and the output reads well. Ramirez called Spark's output "much more readable" than Claude's.

The complaints are consistent too:

• Skills and follow-through. Ramirez says it's "not that good following skills" and he had to nudge it to continue.
• Sandbox friction. He couldn't run Firebase emulators or reach external ports from inside it.
• Efficiency claims. DataCamp measured a net 12% cost increase across three tasks versus Spark 1.2, with uncached input down 19–22%, not 25%. Spark 1.3 was 38% more expensive on the refactor task (DataCamp; our catalog entry).
• Benchmarks. Codus read Meta's three launch charts and found Claude Opus 5 wins all three, including Meta's internal eval.

MakerStack scored the beta 8.0/10, saying persistent async subagents and long-horizon tuning make it "the most interesting Claude Code alternative yet."

Muse Code vs Claude Code

Pick Muse Code if price is the constraint and you're comfortable switching to Standard, or you want auditable multi-agent runs. Pick Claude Code if you need the strongest results on hard tasks today, since Meta's own charts put Opus 5 ahead. Many builders run both: a Claude Code plugin hands reviews and delegated tasks to Muse Code. Our full comparison is in Muse vs Claude.

Frequently asked questions

Is Muse Code free?

The CLI is free to install, but Meta doesn't list a free plan. Paid plans are $5, $15 and $50 a month, or you can sign in with a Meta Model API key and pay per token.

How do I install Muse Code?

On macOS or Linux, run curl -fsSL https://dev.meta.ai/install.sh | sh. On Windows, run irm https://dev.meta.ai/install.ps1 | iex in PowerShell. Then run muse in your project and sign in with your browser or an API key.

Does Muse Code train on my code?

It can. Muse Code starts on the contributor model, where Meta may use your prompts and outputs to improve its products. Run /model and choose muse-spark-1.3 to use the Standard model, which Meta says is not used for training.

Is Muse Code available outside the US?

Yes. Meta expanded Muse Code and the Meta Model API public preview to developers worldwide. The contributor model is limited to select countries Meta hasn't listed. The Muse personal agent is a separate product limited to the US and Canada.

Is Muse Code better than Claude Code?

On Meta's own launch charts, Claude Opus 5 wins all three. Muse Code is cheaper and has strong multi-agent and audit features, but testers report weaker skill-following. Try it on a real task before switching.

What model does Muse Code use?

Muse Spark 1.3 and 1.2, each in Standard and contributor versions. DataCamp found new sessions set to muse-spark-1.3-contributor. The harness can also run local models through Ollama.