Meta launched Muse Code on August 5, 2026, entering the terminal coding agent race with a product that undercuts every competitor on price. The catch: the cheapest tier requires you to share your code with Meta for training. Here is everything you need to know about what it is, what it costs, and whether it belongs in your workflow.
What is Muse Code?
Muse Code is a terminal-based AI coding agent built by Meta Superintelligence Labs. It runs in your terminal on macOS and Linux, takes on complex software engineering tasks across large repositories, and can plan changes, write code, and validate results autonomously.
Unlike Cursor or Windsurf, Muse Code has no GUI, no VS Code extension, and no JetBrains plugin. It is a pure terminal agent, similar in form factor to Claude Code and Kiro CLI.
The underlying model is Muse Spark 1.2, Meta’s coding-focused model that was co-trained specifically with the Muse Code agent harness. This means the model is tuned for long-horizon, repository-scale work rather than one-shot completions.
How to install
Installation is a single command:
curl -fsSL https://dev.meta.ai/install.sh | bashThis works on macOS and Linux. Windows is not currently supported. After installation, you authenticate with your Meta developer account and select your pricing tier.
Pricing: The two tiers explained
Muse Code offers two pricing tiers, and the difference between them is dramatic.
Standard tier
| Token type | Price per million |
|---|---|
| Input | $1.25 |
| Output | $4.25 |
The standard tier is straightforward pay-per-token pricing with no data-sharing requirements. Your code, prompts, and completions remain private and are not used for training future Meta models.
Contributor tier
| Token type | Price per million |
|---|---|
| Input | $0.10 |
| Output | $0.20 |
The contributor tier is roughly 21x cheaper than standard pricing. In exchange, you grant Meta permission to train future models on your prompts, code, and completions.
The data tradeoff
The 21x price difference is not subtle. For a developer spending $100/month on the standard tier, the contributor tier would cost under $5 for the same usage.
But the tradeoff is real. Meta explicitly states that contributor-tier data feeds into training future Meta models. If you work on proprietary code, trade secrets, or anything covered by an NDA, this tier is likely off-limits. For open-source work, personal projects, or code you are comfortable sharing, the economics are compelling.
For context, Claude Sonnet 5 charges $2/$10 per million tokens (introductory, through August 2026) and does not train on your data by default. The Muse Code contributor tier is 50x cheaper on output tokens than Sonnet 5.
Core capabilities
Persistent subagents
Muse Code does not spawn new agents for each subtask. Instead, it runs persistent background agents that remain active throughout a session. These agents carry context, avoid redundant work, and decide when to report back to the main agent. This reduces latency and the constant re-steering that plagues other agent tools on multi-step tasks.
Crash-safe runtime
Muse Code uses a local event log where every model call, tool run, approval, and edit is appended. This makes the runtime replay-exact and restart-safe. If your machine crashes or the process is killed, the agent resumes precisely where it stopped. For long-running tasks, this is a significant advantage over agents that lose all progress on interruption.
Bundled skills
Muse Code ships with built-in slash commands:
/planturns a task into an approval-gated plan/grillstress-tests that plan until it holds up/goalworks toward successful completion of a specified objective
1000+ tool calls in a single session
Meta demonstrated Muse Code making over 1,000 tool calls in a 24-hour GPU kernel optimization task. The agent wrote, compiled, profiled, and progressively improved kernel performance for NVIDIA Hopper GPUs, achieving substantial speedups over baseline implementations.
Multimodal input
Muse Code accepts video and image input. In Meta’s demos, the agent interpreted a fly-through video of a home (an MP4 file piped into the terminal) and produced a visually rich marketing page from it.
Benchmark performance
Muse Spark 1.2 posts the following scores (vendor-reported, not independently verified):
| Benchmark | Muse Spark 1.2 | Claude Opus 5 | GPT-5.6 Terra |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.9% | 86.7% | Lower |
| DeepSWE 1.1 | 59.3% | Higher | Higher |
On Meta’s own benchmark charts, Muse Spark 1.2 trails Claude Opus 5 on every coding benchmark shown, while beating GPT-5.6 Terra and Gemini 3.6 Flash on most. The 82.9% Terminal-Bench score represents a 6.7-point improvement over Muse Spark 1.1.
These are vendor-reported numbers. No independent verification exists yet.
How it compares to Claude Code and Kiro
| Feature | Muse Code | Claude Code | Kiro CLI |
|---|---|---|---|
| Platform | macOS, Linux | macOS, Linux, Windows | macOS, Linux, Windows |
| Cheapest output | $0.20/M (contributor) | $10/M (Sonnet 5) | Model-dependent |
| Data sharing | Yes (contributor tier) | No | No |
| Persistence | Crash-safe event log | Session-based | Session-based |
| Subagents | Persistent background | Per-task | Per-task |
| Open source | No (closed model) | No (closed model) | Kiro Crew is open-source |
The key differentiators for Muse Code are price (at the contributor tier) and the crash-safe runtime. Claude Code has a more mature ecosystem and higher benchmark scores. Kiro Crew offers 24/7 persistent operation and open-source transparency.
Limitations
Beta status. Muse Code is in beta. Expect rough edges, missing features, and potential breaking changes.
No Windows support. macOS and Linux only. Enterprise teams with mixed developer fleets will hit this wall immediately.
Closed model. Unlike Meta’s Llama models, Muse Spark 1.2 has no open-weights option. This is a notable shift from Meta’s open-source positioning.
No IDE integration. Terminal only. No VS Code, JetBrains, or Xcode plugins.
Vendor-reported benchmarks only. No independent evaluation has confirmed Meta’s published scores.
Contributor tier privacy. If you use the cheap tier, Meta trains on your code. Period.
Who should use Muse Code?
Use the contributor tier if: You work primarily on open-source projects, personal experiments, or code that is not proprietary. The pricing is unmatched.
Use the standard tier if: You want Meta’s agent capabilities without sharing code. At $1.25/$4.25, it is still cheaper than Claude Sonnet 5 for the same task.
Skip it if: You need Windows support, IDE integration, or independently verified performance guarantees.
Muse Spark 1.2 also powers Meta’s Ray-Ban smart glasses AI, handling real-time scene understanding, Threads integration, and neural handwriting input.
FAQ
Is Muse Code free? No. There is no free tier. The cheapest option is the contributor tier at $0.10 input / $0.20 output per million tokens, but it requires data sharing.
Does Meta train on my code? Only if you use the contributor tier. The standard tier does not share your data for training.
Can I use Muse Code on Windows? Not currently. Only macOS and Linux are supported in the beta.
What model does Muse Code use? Muse Spark 1.2, a coding-focused model co-trained with the Muse Code agent harness.
Is Muse Spark 1.2 open source? No. Unlike Meta’s Llama models, Muse Spark 1.2 is closed-source with no open-weights release.
How does it compare to Claude Code on benchmarks? Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1 vs Claude Opus 5 at 86.7%. It trails on every published benchmark but costs significantly less.
What is the context window? Meta has not published an exact context window size for Muse Spark 1.2 at the time of writing.
Can it handle large repositories? Yes. Muse Code was specifically designed for repository-scale work, with persistent subagents and context compaction for long-horizon tasks.