Meta launches Muse Code, a coding agent for large repos

Meta has entered the coding-agent race with Muse Code, a terminal tool now in beta that installs with a single command. It runs on the company's Muse Spark model and is built for work inside large repositories, not single-file edits. Mark Zuckerberg says it can handle complete engineering tasks, from planning a change to writing the code and validating the result. Its main design choice is parallelism: rather than one agent working step by step, Muse Code spins up several sub-agents that run at the same time in separate worktrees, so their changes do not collide in your working directory. Meta showed it building six features for a game at once without conflicts.

The pitch is cost. Meta's AI chief, Alexandr Wang, framed Muse Code as a strong option against OpenAI's Codex and Anthropic's Claude Code, especially on price, though the company did not publish pricing or benchmark numbers. That leaves the most important comparison, how good the output actually is, unanswered for now.

Why it matters

If you already pay for Claude Code or Codex, Muse Code adds a third serious option and some pricing pressure, but without published benchmarks you cannot judge quality yet. The parallel-worktree design is the part worth testing on your own repo, since it targets the real bottleneck of running one agent at a time.