MENLO PARK, Calif. — Meta on Wednesday released Muse Code, its first publicly available AI coding agent, entering a fast-growing market led by tools from Anthropic, OpenAI and other artificial intelligence companies.
The terminal-based agent is available in beta and is powered by Muse Spark 1.2, a coding-focused update to Meta’s newest model family. Meta said Muse Code can take on complete software engineering assignments across large repositories, including planning changes, writing code and validating the results.
The release expands Meta’s push to make its AI models useful beyond the company’s consumer apps and social platforms. It also gives developers a new option in a category that has moved rapidly from code completion toward systems capable of handling multistep projects with limited supervision.
Persistent agents and long-running tasks
Muse Code uses a main agent supported by multiple background agents that remain active during a development session. According to Meta’s product announcement, those background agents can gather information, carry out next steps and report back to the primary agent without being recreated for each individual task.
Meta said the persistent design is intended to reduce repeated work and improve performance on difficult, multistep assignments. The agent also maintains a local event log of model calls, tool use, approvals and code changes. That log is designed to let a task resume after a crash without losing its previous state.
The beta includes several built-in workflow commands. One creates an approval-gated plan, another tests the plan for weaknesses, and a third allows the agent to continue working toward a defined goal. Meta said the system can coordinate multiple subagents while keeping their changes separate from a developer’s working copy.
Muse Code is available for macOS and Linux. Meta has not announced a Windows version or provided a timeline for moving the product out of beta.
A coding-focused model update
Muse Spark 1.2 follows the release of Muse Spark 1.1 in July. Meta said the updated model improves code generation, complex debugging, codebase understanding and end-to-end development workflows. The company said it increased the model’s training compute for coding tasks and broadened the range of training environments.
Meta also said it trained Muse Spark 1.2 and Muse Code together so the model would work effectively with the agent’s planning, context-management and subagent features. The model was trained on long-duration tasks that included generating entire repositories, completing large projects and conducting automated research.
In one test described by the company, Muse Code worked through more than 1,000 tool calls over a period of as long as 24 hours while optimizing software used with Nvidia graphics processors. Meta reported performance improvements over its baseline implementations, though the results were produced by the company and have not been independently verified.
Muse Spark 1.2 is also available through the Meta Model API, the company said.
A more crowded coding-agent market
The launch places Meta in more direct competition with Anthropic’s Claude Code and OpenAI’s Codex, both of which are designed to work across software projects rather than respond only to isolated coding prompts.
Meta is positioning Muse Code around repository-scale execution, persistent context and parallel agent work. The company did not disclose how many developers are participating in the beta or provide adoption targets.
The product is the latest release from Meta Superintelligence Labs, the AI organization led by Alexandr Wang. Meta has been expanding its AI product line while investing heavily in models, data centers and computing infrastructure.
For developers and technology buyers, the release adds another major provider to an increasingly competitive field. The next test will be how Muse Code performs in real production environments, where reliability, security controls and the ability to recover from errors may matter as much as benchmark scores.