Muse Code

Muse Code

Terminal coding agent by Meta AI, powered by Muse Spark 1.2, enabling persistent background agents, repository-scale execution, and built-in verification for efficient software development.

terminal coding agentmeta aimuse sparkbackground agentsrepository-scale execution
Open site
image for Muse CodeOpen site

What is Muse Code?

Muse Code (beta) is a terminal coding agent developed by Meta AI, powered by the Muse Spark 1.2 model. It runs on macOS and Linux and handles complex software engineering tasks across large repositories—planning changes, writing code, and validating results. The tool coordinates multiple persistent subagents for each task, enabling faster, more accurate problem-solving with less manual intervention. It's built for developers who need an autonomous assistant that can work through long, multi-step coding challenges directly from the command line.

Application scenarios

  • Large repository development

    Tackle complex engineering tasks that span entire codebases, from planning to implementation.

  • Long-running autonomous tasks

    Let background agents work continuously without constant oversight, reducing latency on difficult, multi-step jobs.

  • Code debugging and validation

    Use built-in verification to stress-test plans and ensure changes hold up before completion.

  • End-to-end project generation

    Build whole repositories and large-scale projects with the model's long-horizon training.

  • Multimedia processing

    Input video files (e.g., a home tour) and generate rich, functional outputs like marketing and booking pages.

  • Crash recovery workflows

    Resume work precisely after interruptions, thanks to a replay-exact runtime design.

Core Features

  • Async background agents

    Persistent subagents stay active throughout a session, carrying out next steps and deciding when to report back—reducing redundant work and manual steering.

  • Replay-exact runtime

    A local event log records every model call, tool run, approval, and edit, making the agent restart-safe after crashes.

  • /plan skill

    Converts a task into an approval-gated plan, so you review and approve before execution.

  • /grill skill

    Stress-tests a plan until it holds up, catching weaknesses before you commit to a path.

  • /goal skill

    Works toward successful completion of a specified objective, keeping the agent on track.

  • Repository-scale execution

    Handles tasks across large codebases, not just isolated files.

  • Muse Spark 1.2 model

    Improved code generation, complex debugging, codebase understanding, and end-to-end developer workflows.

  • Co-training with Muse Code

    The model was trained alongside the tool to maximize harness compatibility and coding usability.

Target users

Muse Code suits professional software engineers and development teams working on large, complex repositories. It's also valuable for technical leads who want to delegate multi-step coding tasks to an autonomous agent, and for developers handling long-horizon projects like whole-repository generation or auto-research. Anyone comfortable with a terminal-based workflow and needing persistent, low-intervention coding assistance will find it useful.

How to use Muse Code?

Installation is straightforward: run the provided curl command (curl -fsSL https://dev.meta.ai/install.sh | bash) on macOS or Linux. Once installed, you interact with Muse Code through the terminal, issuing tasks and using built-in skills like /plan, /grill, and /goal to structure and validate your work. For full details and updates, check the official Meta AI research page.

Effect review

Muse Code delivers a genuinely capable autonomous coding experience, especially for developers managing large, complex repositories. The persistent background agents and replay-exact runtime address two real pain points: latency on multi-step tasks and vulnerability to crashes. The bundled skills—planning, stress-testing, and goal execution—provide a structured workflow that reduces the need for constant human oversight. The co-training with Muse Spark 1.2 shows a thoughtful approach to model-tool integration, and the ability to handle long-horizon tasks like whole-repository generation sets it apart from simpler coding assistants. While it's still in beta and limited to macOS and Linux, the feature set suggests strong practical value for teams that need a reliable, low-intervention coding partner.

Frequently Asked Questions

What is Muse Code?
Muse Code is a terminal-based coding agent by Meta AI, powered by Muse Spark 1.2, that enables persistent background agents, repository-scale execution, and built-in verification for efficient software development.
How does Muse Code handle large codebases?
Muse Code supports repository-scale execution, allowing it to analyze, modify, and test code across an entire repository, not just individual files.
Can Muse Code run tasks in the background?
Yes, Muse Code supports persistent background agents, enabling it to continue working on tasks asynchronously while you focus on other work.
Does Muse Code include automated testing or verification?
Yes, Muse Code has built-in verification capabilities, automatically running tests and checks to validate code changes before completion.
Is Muse Code compatible with existing development workflows?
Yes, Muse Code operates from the terminal, integrating seamlessly with standard Git workflows, command-line tools, and CI/CD pipelines.
What makes Muse Code different from other coding assistants?
Muse Code combines persistent background execution, repository-wide scope, and built-in verification in a single terminal agent, offering a more autonomous and scalable approach to coding assistance.

Muse Code - AI Tool Detail

Terminal coding agent by Meta AI, powered by Muse Spark 1.2, enabling persistent background agents, repository-scale execution, and built-in verification for efficient software development.

Category：Programming Assistant

Visit Link：https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2

Tags：terminal coding agent、meta ai、muse spark、background agents、repository-scale execution