Meta debuts Muse Code agent

Meta launched Muse Code on August 5 alongside Muse Spark 1.2, giving the company its first dedicated coding agent as it pushes deeper into developer tools. The new product is aimed at software engineering tasks across large code bases and is being positioned against Anthropic’s Claude Code and OpenAI’s Codex. Meta said Muse Code is built to run in the terminal and work through repository tasks with an auditable record of what it did. Reuters, CNBC and Yahoo Finance each reported the launch on Wednesday. ### Why did Meta pair Muse Code with Muse Spark 1.2? Meta said Muse Spark 1.2 is the model that powers Muse Code, making the launch a combined model-and-agent release rather than a standalone app debut. Yahoo Finance reported that Muse Spark 1.2 is a coding-focused update to Muse Spark, which Meta rolled out last month, and described Muse Code as the coding harness that runs the model. (tech.yahoo.com) Reuters reported that Muse Code is designed to help developers write and debug software. TechCrunch and Engadget described it as a terminal-based coding agent, while 9to5Mac said the beta is available for macOS and Linux. ### What exactly is Meta selling to developers? Muse Code is a coding agent that Meta says can take on larger software tasks rather than just answer code questions inline. (finance.yahoo.com) TechCrunch reported that Meta Chief Executive Mark Zuckerberg said bigger jobs can be split into parallel sub-agents working in isolated worktrees, so the user’s main working copy is not touched. (tech.yahoo.com) Unite.ai, ADTmag and other coverage said the product is built for longer-running jobs and can resume work after interruptions. ADTmag reported that Meta said Muse Code keeps a local event log of model calls, tool activity, approvals and edits, allowing tasks to restart after a crash instead of beginning again. (techcrunch.com) ### Why are audit logs getting so much attention? Seeking Alpha and follow-on coverage highlighted audit logs as one of the product’s defining features. The logs cover repository tasks, giving developers a replayable record of what the agent changed and which steps it took. ADTmag reported that Meta tested Muse Code on a graphics-processor optimization job involving more than 1,000 tool calls and runs lasting up to 24 hours. (unite.ai) That emphasis puts Meta’s pitch on operational controls as much as raw model performance. CNBC reported that Meta is ramping up investment in AI models and services as it tries to compete with Anthropic and OpenAI. ### How is Meta trying to compete with Anthropic and OpenAI? Business Insider, TechCrunch and CNBC reported that Meta is presenting Muse Code as a lower-cost alternative to Anthropic’s Claude Code and OpenAI’s Codex. (adtmag.com) Reuters said the launch comes as large AI companies race to monetize coding assistants, one of the most commercially active uses of generative AI. (cnbc.com) Yahoo Finance reported that Meta stock was flat in after-hours trading following the announcement. The same report said Meta compared Muse Spark 1.2 with models including Anthropic’s Opus 5, GPT 5.6 Terra, Gemini 3.6 Flash and Grok 4.5, even while describing the release as short of the very top frontier tier. (tech.yahoo.com) ### What should buyers and developers watch next? The beta release is available now through Meta’s developer tooling, with Muse Code tied to Muse Spark 1.2 and Meta’s broader model API rollout. 9to5Mac reported initial support for macOS and Linux, while release-tracking coverage said Meta is expanding access globally through Muse Code and the Meta Model API. (finance.yahoo.com) Anthropic, OpenAI and Google already have products in the same category, and Meta’s next test will be whether developers adopt Muse Code for sustained repository work rather than one-off code generation. For now, the immediate next step is the beta period, where Meta is asking developers to use the terminal-based agent and evaluate its long-running task support, logging and pricing against rival tools. (9to5mac.com) (cnbc.com)