Meta has launched Muse Code in beta, introducing a new terminal coding agent designed to compete with developer tools including Anthropic’s Claude Code and OpenAI Codex.
Powered by the newly released Muse Spark 1.2 model, Muse Code can plan software changes, write code and validate its results across large repositories. The agent can also coordinate multiple persistent subagents to complete difficult tasks with less user intervention.
The background agents remain active throughout an entire session rather than being created separately for individual requests. They can gather information, carry out additional steps and decide when to communicate their findings to the main agent.
Meta said the persistent structure reduces repeated information gathering and lowers latency during complex tasks involving multiple stages.
Muse Code records every model request, tool execution, approval and code edit in a local event log. This allows the runtime to recreate a session exactly and resume from the same point following a crash or interruption.
The system is intended to support long running software development tasks without losing previous work or requiring the user to restart the process.
Muse Code also includes several built in skills. The /plan command converts a task into a plan that requires user approval. The /grill command tests that plan for potential weaknesses, while /goal directs the agent toward completing a specified objective.
In one example shared by Meta, a user provided Muse Code with a video tour of a home. The agent interpreted the video and created a vacation property marketing and booking website based on its contents.
See what prediction markets are pricing into the day's biggest stories.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Muse Spark 1.2 is a coding focused update to Muse Spark 1.1. Meta said the model improves code generation, complex debugging, codebase understanding and complete developer workflows while maintaining broader agent capabilities.
The company increased the amount of training compute dedicated to coding tasks and expanded the variety of environments used during training.
Meta also trained Muse Spark 1.2 alongside Muse Code to improve how the model works with the agent’s tools, planning system, context management and subagents.
The training process included long horizon coding tasks such as creating entire repositories, completing large projects and conducting automated research. The model uses planning to organize its work and context compaction to retain important information during extended sessions.
Meta also used Muse Spark 1.1 to create challenging coding environments and instruction following templates. The earlier model then evaluated candidate solutions, producing additional training data for Muse Spark 1.2.
During one test, Muse Spark 1.2 used Muse Code to optimize GPU kernels through more than 1,000 tool calls over sessions lasting as long as 24 hours.
The agent repeatedly wrote, compiled, profiled and improved code for Nvidia Hopper GPUs. Meta said the model produced substantial performance gains compared with the provided baseline implementations.
Muse Code is available in beta on macOS and Linux. Muse Spark 1.2 is also available through the Meta Model API with expanded global access.
Meta said it plans to introduce additional Muse Code features and more capable models as it expands its presence in the increasingly competitive AI coding market.