Muse Code Review (2026): Meta's Terminal AI Coding Agent

Overview

Muse Code is Meta Superintelligence Labs' first terminal AI coding agent, released in beta on August 5, 2026 and powered by the company's Muse Spark 1.2 model. It positions itself in the increasingly crowded category of agentic command-line coding tools, alongside Claude Code, OpenAI's coding agent, Cursor, and Kiro. Rather than acting as an autocomplete assistant, Muse Code takes on repository-scale software engineering: it plans changes, writes code, and validates the results, coordinating multiple persistent sub-agents per task.

This review is based on Meta's official announcement and reputable reporting from launch week. Because Muse Code is a beta product, some figures and behaviors are provisional and may change; where numbers are reported rather than officially fixed, they are hedged accordingly.

At a Glance

Pricing snapshot

The two-tier structure is the headline commercial story. The Standard tier is priced aggressively for a frontier coding agent and keeps your data out of training. The Contributor tier drops token costs by roughly an order of magnitude, but the discount is explicitly a trade: Meta may use your prompts and completions to improve future models. Reporting also indicates the Contributor tier is rate-limited on a rolling five-hour window and available only in select countries. Both sets of numbers should be treated as beta pricing that could shift.

Key Features

Persistent async background sub-agents. Muse Code's defining design choice is its agent architecture. It runs a main agent loop augmented by a set of async background sub-agents that stay active throughout a session rather than being spawned and discarded per task. According to Meta, this persistence avoids redundant information gathering, lets the sub-agents carry out next steps and choose when to report back, and reduces both latency and the need for hand-holding on difficult, multi-step work.

Plan, write, and validate. The agent is built around a full task loop rather than single edits. Bundled skills make this explicit: /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal works toward completion of a stated objective. This structure gives developers a checkpoint before the agent starts making sweeping changes.

Crash-safe event-log runtime. Muse Code appends every model call, tool run, approval, and edit to a local event log. Meta describes this as a single source of truth that makes the runtime replay-exact and restart-safe, so after a crash the agent can resume precisely where it stopped. For long-running, unattended tasks, this is a meaningful reliability feature.

Muse Spark 1.2. The underlying model is a coding-focused update to Muse Spark 1.1, co-trained with the Muse Code harness for maximum compatibility. Meta says it scaled up coding training compute and widened environment diversity, and highlighted a case study in which the model iteratively optimized GPU kernels over more than 1,000 tool calls and up to 24 hours of continuous work. As with any vendor benchmark, these results are self-reported and best read as directional.

Strengths and Limitations

Muse Code's strengths center on autonomy and endurance. The persistent sub-agent model and crash-safe runtime are genuinely aimed at long-horizon tasks that other tools struggle to sustain, and the Standard-tier pricing undercuts several established rivals. Backing from Meta's compute and the availability of Muse Spark 1.2 through the Meta Model API add credibility.

The limitations are equally clear. As a beta, stability, features, and pricing may change without notice. It supports only macOS and Linux, leaving Windows developers without a native client, and it is terminal-only, with no GUI, mobile app, or browser extension. The most consequential caveat is the Contributor tier: the large discount is contingent on allowing Meta to train on your prompts and completions, which is a poor fit for proprietary or sensitive codebases, and that tier also carries rate limits and regional restrictions. Teams that value confidentiality should budget for the Standard tier.

Pros and Cons

Pros

• Persistent async background sub-agents reduce latency and steering on multi-step tasks
• Crash-safe event-log runtime resumes long jobs exactly where they stopped
• Purpose-built Muse Spark 1.2 model co-trained with the coding harness
• Standard tier is aggressively priced versus rivals
• Bundled /plan, /grill, and /goal skills structure and stress-test work

Cons

• Contributor tier's discount requires letting Meta train on your data
• Early beta; stability and features may change
• macOS and Linux only, no native Windows support
• Terminal-only, with no GUI or mobile app
• No free plan or trial; billing is purely usage-based

Who It's For

Muse Code is best suited to professional software and ML engineers who live in the terminal, work on large codebases, and want an autonomous agent for planning, refactoring, debugging, and long-running tasks. Cost-sensitive individual developers may find the Contributor tier attractive, provided they are comfortable with the data-sharing terms and work on non-sensitive code. Teams handling proprietary or regulated codebases should default to the Standard tier or evaluate carefully before adopting a beta tool.

Verdict

Muse Code is a credible and technically interesting entry into the terminal coding-agent race. Its persistent sub-agents and crash-safe runtime address real weaknesses in long-horizon agentic work, and its Standard pricing is competitive. But it is still a beta limited to macOS and Linux, and its most eye-catching price point comes with a data-for-discount tradeoff that many professional teams will decline. For developers already invested in Claude Code or Cursor, Muse Code is worth watching and testing rather than switching to outright, at least until it matures beyond beta.

FAQ

Is Muse Code free? No. There is no free plan or trial; both tiers are usage-based on token pricing.

Does Meta train on my code? Only on the Contributor tier, where Meta may use your prompts and completions to train models. The Standard tier states data is not used for training.

What platforms are supported? macOS and Linux, via a single terminal install command. There is no native Windows client at launch.

What model does it use? Muse Spark 1.2, which is also available through the Meta Model API.