Meta Muse Code enters the AI coding-agent market in beta
Meta has released Muse Code, a terminal AI agent for large codebases that can plan, write and validate software changes.
By Theo Nakamura · Staff Writer
· 3 min read
Meta Muse Code is now in beta, giving developers a new terminal-based AI tool intended to handle work across large software repositories. For investors tracking Meta’s AI push, the launch puts the company in the same product category as OpenAI’s Codex and Anthropic’s Claude Code, though its real-world performance against those tools remains to be seen.
Meta released Muse Code on Aug. 5 for macOS and Linux, according to 9to5Mac and VentureBeat. It runs in the terminal, the text-command interface used by developers, rather than through a standalone graphical app. The tool can be installed with a one-line command, 9to5Mac reported.
Chief Executive Mark Zuckerberg said in a post on X that the agent is designed to take on full software-engineering assignments across large repositories, or repos. Meta says those assignments can include mapping out changes, generating code and checking whether the result works.
What does Meta Muse Code do for large codebases?
Muse Code is powered by Muse Spark 1.2, a coding-focused update to Meta’s Muse Spark model family, according to VentureBeat and 9to5Mac. Meta describes the model as improved for code generation, difficult debugging and understanding an existing codebase.
Its approach to larger tasks relies on multiple agents, which are AI processes assigned pieces of a broader job. VentureBeat reported that Meta keeps specialized background agents running throughout a coding session, rather than creating a new one for each request. The company’s aim is to avoid repeatedly gathering the same context about a project.
For sufficiently large jobs, Meta says Muse Code can split work among parallel sub-agents. Each works in an isolated Git worktree, a separate copy of a code project that lets changes be made without altering a developer’s primary working copy. Zuckerberg said Meta tested the system by having it create six game features at once without collisions. That was a company testing claim, not an independently verified result.
Why the design may matter to developers
Long-running coding tasks can be disrupted if a tool fails or a session ends. Meta says Muse Code maintains a local event log recording model calls, tools used, approvals and edits. According to 9to5Mac, that record is intended to let work resume after a crash. VentureBeat reported Meta characterizes the design as restart-safe and auditable.
The feature set is relevant to teams maintaining established software, where a change may touch many files and require testing against existing code. It does not remove the need for developer review: Meta’s stated capabilities cover producing and validating changes, while the reports do not establish how reliably the beta tool performs across outside projects.
Muse Code is Meta’s entry into a crowded market for coding agents. The immediate questions for developers will be whether its persistent agents, parallel worktrees and event log save time in daily work, and how the beta compares in practice with Codex and Claude Code.
This story draws on original reporting from TechCrunch.
