Meta has released Muse Code, a beta terminal-based coding agent designed to demonstrate the capabilities of its latest large language model, Muse Spark 1.2. The tool positions Meta as a direct rival to established coding agents such as OpenAI Codex and Anthropic's Claude Code, and is intended to help developers plan codebase changes, write code and validate the results. It marks Meta's continued push into AI-assisted software development, though notably the underlying model remains closed-weight and cloud-hosted, a departure from the company's previously open approach with its Llama models.

Muse Code works as an agent orchestrator, spinning up background agents that maintain a shared context file while sub-agents work in isolated work trees, logging actions before execution so no work is lost. CEO Mark Zuckerberg cited tests in which the platform built six features for a game simultaneously without collisions, and another where agents made over 1,000 tool calls during a 24-hour kernel optimisation task. Muse Spark 1.2 is the third model release in four months from Meta Superintelligence Labs, formed in June 2025, and addresses coding weaknesses acknowledged when the multi-modal model first launched in April; Zuckerberg has not ruled out eventually open-sourcing Muse Spark.