Meta launches Muse Code, a terminal AI agent

Meta released Muse Code in beta on August 5, a terminal coding agent for macOS and Linux, installable with a single command. It runs on Muse Spark 1.2, an update to Meta’s coding model with gains in generation, debugging, codebase understanding, and long-running workflows. CEO Mark Zuckerberg described the agent’s reach in a social post:

Complete software engineering tasks across large repos: planning changes, writing code, validating the results.

The distinguishing feature is how it scales a large job. Zuckerberg explained that once a task is sizable, the agent splits the work across sub-agents that run at the same time, each confined to its own isolated copy of the repository so the developer’s actual working files stay untouched until changes are reviewed and merged. Muse Code also runs persistent background agents that stay active for a whole session, building up project context instead of restarting from scratch on each task, and ships default skills including /plan for turning a request into an approval-gated plan and /grill for stress-testing that plan before execution. Meta is offering a low-cost contributor tier for users who opt into data sharing.

What it means for you

Meta is now a fourth serious entrant in the terminal coding-agent category alongside Claude Code, OpenAI’s Codex, and Cursor, at a moment when picking a coding agent already means comparing on more than raw benchmark scores and picking a desktop AI agent generally comes down to a handful of concrete tradeoffs. The parallel-subagent worktree model and the discounted data-sharing tier are the two details worth testing directly: the first for whether it actually reduces your review burden on large changes, the second for whether the price break is worth what you give up on the code you paste in.