Meta launches Muse Code, an AI agent for large code bases

Meta Unveils Muse Code: A Powerful AI Agent for Large Software Projects

Meta Platforms has made a significant move into the competitive landscape of AI-powered software development with the beta launch of Muse Code. Announced on August 5, 2026, this new terminal-based AI coding agent is designed to tackle "complete" and complex software engineering tasks across extensive code repositories.

The introduction of Muse Code marks Meta's most serious entry yet into the realm of agentic coding, positioning it as a direct competitor to existing tools like Anthropic's Claude Code and OpenAI's Codex. Unlike some of Meta's previous open-source AI efforts, such as the Llama model family, Muse Code is a proprietary offering, built to address the growing demand for AI systems that can do more than just suggest code – they can plan, execute, and validate entire software changes.

What is Muse Code and Why Does it Matter?

At its core, Muse Code is an AI agent built to streamline and automate complex software development workflows. It's powered by Muse Spark 1.2, a newly updated version of Meta's foundational AI model specifically optimized for coding tasks. The agent aims to push AI deeper into software engineering by enabling systems to inspect projects, make changes, test their work, and sustain these efforts over hours, even days.

The significance of Muse Code lies in its ambition to handle "long-horizon" tasks – multi-step, intricate assignments that typically require extensive human developer interaction. Traditionally, AI coding assistants have excelled at generating snippets or offering suggestions. Muse Code, however, is designed to take on the entire lifecycle of a software change, from understanding the initial request to validating the final results across large, enterprise-grade codebases.

How Muse Code Works: An Agentic Approach

Muse Code distinguishes itself through a sophisticated architecture that employs a system of coordinated AI agents. Instead of running a single, isolated AI process for each subtask, Muse Code leverages:

• Persistent Background Agents: These specialized agents remain active throughout an entire coding session, continuously gathering information and carrying out supporting tasks. This approach allows the AI to build and maintain context over time, reducing redundant investigation and the amount of "steering" a developer needs to provide for multi-step tasks.
• Parallel Sub-agents: For larger, more complex assignments, Muse Code can delegate parts of the task to separate sub-agents. These sub-agents work in parallel within isolated "worktrees," ensuring that the developer's main working copy of the code remains untouched. This capability allows for simultaneous feature development without collisions, as demonstrated by Meta's internal testing where it built six game features concurrently.
• Local Event Log for Auditability and Resilience: Every action taken by Muse Code – including model calls, tool usage, developer approvals, and code edits – is recorded in an append-only local event log. This detailed record makes the agent's runtime "replay-exact" and "restart-safe." If a task encounters a crash or interruption, the agent can resume precisely where it left off, a crucial feature for tasks that can span many hours.

The agent is accessible via a simple one-line command in the terminal for macOS and Linux users, signifying its developer-centric, command-line interface rather than a graphical application.

Key Capabilities and Features

Muse Code, powered by Muse Spark 1.2, brings several advanced capabilities to the table:

• End-to-End Software Engineering Tasks: It can plan changes, write new code, and validate the results across large repositories, aiming to automate complete software engineering workflows.
• Enhanced Code Generation and Debugging: Muse Spark 1.2, the underlying model, boasts improvements in code generation, the ability to diagnose and fix complex bugs, and a deeper understanding of large codebases.
• Specialized Commands: The agent includes built-in commands like /plan to generate an approval-gated plan for a task, /grill to stress-test that plan, and /goal to work towards a specified objective.
• Long-Running Task Management: Meta has tested Muse Code on demanding tasks, such as graphics-processor optimization, involving over 1,000 tool calls and running for up to 24 hours. This highlights its capacity for sustained, complex work.
• Codebase Understanding Beyond Generation: Beyond just writing code, Muse Code leverages AI agents to map and retrieve internal knowledge across large-scale data systems. This helps developers understand existing systems, service relationships, dependencies, and historical changes, potentially cutting investigation time significantly.

Pricing and Accessibility

Muse Code is available with a pay-as-you-go pricing model through the Meta Model API. The standard rates are $1.25 per million input tokens and $4.25 per million output tokens.

In a strategic move to attract developers and gather feedback, Meta is also offering a "contributor tier." This tier provides significantly reduced rates – $0.10 per million input tokens and $0.20 per million output tokens (or $0.30 per million total tokens, depending on the source) – in exchange for users agreeing to provide feedback that will be used to train and improve the coding agent. It's important for enterprises with proprietary codebases to note that the default on-ramp for Muse Code sends developers' code and prompts into Meta's training pipeline, a practice that requires conscious opt-out by moving to standard pricing.

Industry Implications and Meta's AI Strategy

The launch of Muse Code signifies Meta's heightened ambition in the AI developer tools space. While the company has previously contributed to open-source AI with models like Llama and its code-specific variant, Code Llama, Muse Code represents a proprietary product designed for direct commercial deployment.

This move places Meta in direct competition with other tech giants and startups that have been rapidly advancing agentic coding solutions. The focus on long-horizon tasks and persistent agents suggests a shift in the AI coding paradigm – from mere assistance to more autonomous, goal-oriented software engineering. The ability to manage complex tasks across large repositories could significantly boost developer productivity and lower the barrier to entry for managing intricate systems.

Muse Code highlights Meta's broader strategy to integrate powerful AI capabilities across its ecosystem and offer advanced tools to developers. As AI continues to evolve, agents like Muse Code are poised to redefine how software is built, maintained, and understood, pushing the boundaries of what automated systems can achieve in complex engineering environments.

Frequently Asked Questions

What is Meta Muse Code?

Meta Muse Code is a new AI coding agent launched in beta by Meta Platforms. It's a terminal-based tool for macOS and Linux designed to handle complex software engineering tasks, including planning, writing, and validating code changes across large code repositories.

What AI model powers Muse Code?

Muse Code is powered by Muse Spark 1.2, a coding-focused update to Meta's proprietary AI model. This model has been specifically trained to excel in code generation, debugging, and understanding large codebases.

How does Muse Code handle complex or long-running tasks?

Muse Code uses a unique architecture involving persistent background agents that maintain context throughout a session and parallel sub-agents that can work on different parts of a task simultaneously in isolated environments. It also features a local event log that allows tasks to be resumed after interruptions, making it suitable for jobs lasting many hours.

Is Muse Code open source, and what are its pricing options?

Unlike some of Meta's other AI models like Code Llama, Muse Code is a proprietary, cloud-only offering. It operates on a pay-as-you-go model with standard pricing of $1.25 per million input tokens and $4.25 per million output tokens. Meta also offers a "contributor tier" at a reduced rate ($0.10/$0.20 per million tokens) for users who agree to share their usage data for model training.