Meta Launches Muse Code to Take On Claude Code and OpenAI Codex

Meta has launched Muse Code in beta, a competitor to apps like OpenAI Codex and Anthropic's Claude. Available for macOS and Linux, Muse Code is built to handle end-to-end engineering workflows, including planning code changes, writing code, debugging and validating results.

Unlike conventional AI coding assistants that respond to one prompt at a time, Muse Code employs multiple persistent background agents that continuously work alongside the primary agent, enabling it to solve long-running, multi-step software development tasks with minimal user intervention.

The coding agent also features a replay-safe runtime architecture that logs every model interaction, tool execution and code edit. If interrupted, developers can resume tasks exactly where they stopped, allowing the system to work on projects that span several hours or even days.

"It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update," Mark Zuckerberg, Meta CEO, said.

Muse Code includes built-in workflows such as /plan, which converts requests into structured development plans, /grill, which stress-tests those plans, and /goal, which autonomously works toward completing assigned objectives. In one demonstration, the company showed the tool generating a vacation rental website after analyzing a fly-through video of a property.

The launch is accompanied by Muse Spark 1.2, an upgraded coding-focused AI model trained specifically for code generation, debugging, repository understanding and long-horizon software engineering tasks. Meta said the model was co-trained with Muse Code to optimise performance in agentic coding environments.

"In Muse Spark 1.2, we significantly scaled up training compute on coding tasks while expanding training environment diversity. The model also maintains its strength in other key areas like general agents," the company said.

According to the company, Muse Spark 1.2 outperforms earlier versions across multiple coding benchmarks and has been trained on large-scale repository generation, end-to-end software projects and automated research workflows.

Meta also demonstrated the model optimising NVIDIA Hopper GPU kernels over more than 1,000 tool calls across 24 hours, iteratively improving performance relative to baseline implementations.

Muse Spark 1.2 is now available through Muse Code as well as the Meta Model API with expanded global availability.

Meta said larger and more capable AI models are also in development as it continues to invest in agentic software engineering tools.

Last month, the company introduced Muse Image, its first image generation model from Meta Superintelligence Labs.