Meta has officially announced the launch of Muse Code, a new AI-powered programming tool, currently available in beta for Mac and Linux systems only. With this launch, Meta directly enters the "coding agents" arena, a field currently dominated by Anthropic's Cloud Code and OpenAI's Codex.

The new tool operates entirely within a terminal, without any traditional graphical interface, and is specifically designed to handle complex programming tasks within large repositories, from writing the first line of code to running the final code test.

Muse Code offers Meta's software agent for planning, writing, and testing code within large repositories.

The agent operates without a graphical interface, with backup agents continuously performing tasks throughout the session.

Muse Spark supports 1.2 million codebases and visual contexts, with the ability to handle lengthy programming tasks.

Its launch puts Meta in direct competition with Cloud Code and Codex through competitive pricing.

Muse Code: A Software Agent for Large Repositories.

With Muse Code, Meta aims to provide a comprehensive solution that helps developers manage their large software projects without constantly switching between multiple windows and programs. To achieve this, the agent fully automates three essential tasks: planning the required modifications, writing the code itself, and finally, testing the results for accuracy.

In the background, a set of "background agents" remains active throughout the entire session, rather than creating a new agent for each task as is the case with other tools. This specific design minimizes the need to collect the same information repeatedly, thus reducing the tool's response time during lengthy tasks.

The tool also maintains a detailed execution log that documents each step of the process, significantly reducing the risk of any code modifications remaining incomplete in case of a sudden crash during execution.

The tool provides developers with three direct commands that can be used depending on the task:

the `/plan` command converts the required task into a detailed work plan, which only begins execution after explicit developer approval.

The `/grill` command tests the robustness of this plan, revealing any potential vulnerabilities or weaknesses before actual execution begins.

The `/goal` command ensures the agent continues to work automatically and continuously until it reaches its defined end goal without requiring repeated developer intervention.

The Muse Spark 1.2 engine and processing power

work behind the scenes. A new AI model, Muse Spark 1.2, is responsible for driving all the complex operations performed by the tool. This model was co-trained with Muse Code itself over long-term workflows, including building complete software repositories from scratch, as well as fully integrated engineering projects.

The model proved its practical capabilities in a rigorous test related to optimizing the performance of a GPU core. The test involved more than 1,000 consecutive calls to the software tools, running continuously for 24 hours.