Claude Code, Codex... and now Muse Code

Muse Code combines persistent agents with Muse Spark 1.2 to modify large repositories, resume interrupted sessions, and handle long-running tasks.

Muse Code is a terminal-based coding agent currently available in beta on macOS and Linux. It handles complex changes across multiple files, from planning and code generation to validation.

The system combines a main agent with several asynchronous background agents that remain active throughout the session. They can prepare subsequent steps, continue research, and report their findings without being recreated for each subtask.

A local event log sequentially records model calls, tool use, approvals, and edits. This architecture makes it possible to replay a session exactly and resume work after an interruption, including tasks that run for several hours.

Muse Code also includes several specialized commands. `/plan` creates a plan that requires approval, `/grill` stress-tests that plan, and `/goal` continues working toward a specified objective until completion. One demonstration also shows the system interpreting a video walkthrough of a house and building a marketing and booking website from it.

The agent is powered by Muse Spark 1.2, a coding-focused update co-trained with its execution environment. Meta says it has improved code generation, complex debugging, large-repository understanding, and long-running tasks. In a test focused on optimizing GPU kernels on NVIDIA Hopper hardware, the system completed more than 1,000 tool calls during a session lasting up to 24 hours.

Muse Spark 1.2 is available through Muse Code and the Meta Model API. The announcement does not yet specify access requirements, usage limits, or Windows availability.