Meta launches Muse Code, a terminal coding agent built to rival Claude Code and Codex
Meta released a beta of Muse Code, a terminal-based coding agent powered by its Muse Spark 1.2 model, aimed squarely at the same territory as Claude Code and OpenAI's Codex — planning changes, writing code, running tests, and validating results across an entire repository rather than one file at a time.
What's actually different, not just "another coding agent"
Two design choices stand out. First, when a task is large enough, Muse Code spins up persistent sub-agents that work in parallel using isolated git worktrees — background agents that stay alive across a whole session instead of being spun up and torn down per task. Second, and more practically useful day to day: if Muse Code crashes twenty hours into a long-running task, it resumes exactly where it stopped, rather than losing the work and forcing a re-prompt from scratch.
Anyone who's had a long agent run die near the finish line and had to start over knows this is the actual pain point in long-running AI coding tasks — not raw model quality, but losing hours of progress to an infrastructure hiccup. A tool that treats "resume after crash" as a first-class feature is solving a problem competitors mostly leave to the user.
Pricing and positioning
Muse Code uses the same pay-as-you-go pricing as the underlying Muse Spark model by default — $1.25 per million input tokens and $4.25 per million output tokens — undercutting where Claude Code and Codex typically land. Meta's Chief AI Officer has also mentioned a lower-cost "contributor tier" is coming. Whether Muse Code earns a place next to the incumbents will come down to the usual thing benchmarks can't fully capture: how it actually behaves on a real, messy codebase over a multi-hour session, not a demo.
Source: techcrunch.com