Meta launches Muse Code, a terminal coding agent to challenge Claude Code and Codex

Meta has entered the AI coding-agent market with Muse Code, a terminal-based agent unveiled in beta on August 5, 2026, built on the company's updated Muse Spark 1.2 model.

What's new

According to TechCrunch's report on the announcement, "Muse Code, which is currently available in beta, can accomplish 'complete software engineering tasks across large repos,' Meta CEO Mark Zuckerberg said," with those tasks including "planning changes, writing code, validating the results." The tool "can be installed with a single command" and "is powered by Meta's previously released coding model, Muse Spark."

Muse Code is a command-line agent, not a GUI or IDE plugin. For large jobs, it launches sub-agents that work in parallel inside isolated worktrees so a developer's primary working copy stays untouched, and it maintains persistent background agents plus a local event log of model calls, tool use, approvals, and edits so work can resume after a crash. It runs on Muse Spark 1.2, an update to the Muse Spark 1.1 model Meta introduced in July, at the same pay-as-you-go pricing as the base model: $1.25 per million input tokens and $4.25 per million output tokens on the standard tier, or a discounted $0.10/$0.20 per million tokens on a "contributor tier" that requires users to allow their data to be used for training.

Context

Muse Code is the first coding-specific product to ship from Meta Superintelligence Labs, the unit AI chief Alexandr Wang has been building out since Meta's Muse model family launched earlier this year with Muse Spark, Muse Image, and Muse Video. It arrives roughly a month after Muse Spark 1.1 debuted as Meta's first model under Wang's reorganized AI research effort, and positions Meta as a direct entrant into a coding-agent field currently led by Anthropic's Claude Code, OpenAI's Codex, and, most recently, xAI's Grok Build.

Why it matters

Terminal-native coding agents have become one of the most competitive fronts in applied AI, and Meta has been notably absent from it until now. Muse Code's architecture — parallel sub-agents in isolated worktrees, persistent background sessions, crash-safe replay — mirrors design choices already proven out by Claude Code and Codex, suggesting Meta is prioritizing fast catch-up over novel workflow bets for this first release. The contributor-tier pricing, which trades a steep discount for training-data rights, is also a notable structural choice: it gives Meta a built-in mechanism to collect real coding-agent usage data to improve Muse Spark, the same lever OpenAI and Anthropic have used less explicitly through their own agent products.