Meta drops Muse Code to tackle full repos in the terminal

Coding agents are evolving fast, and Meta definitely does not want OpenAI and Anthropic to own your terminal.

Meta launched Muse Code in beta. It is a new terminal coding agent designed to tackle complete software engineering tasks across massive codebases, powered by Meta's freshly updated Muse Spark 1.2 model.

Parallel sub-agents in isolated worktrees

Unlike traditional autocomplete tools, Muse Code operates directly from your command line to plan changes, write code, and validate results across entire repositories. Installation takes a single command.

Under the hood sits Muse Spark 1.2, a closed-source multimodal model supporting a massive 1 million token context window, accessible via API. But the real novelty is in how Muse Code manages the work.

When handed a large project, the agent does not touch your active local files. Instead, it fans out to multiple parallel sub-agents running in isolated git worktrees. Meta CEO Mark Zuckerberg noted that during testing, the agent built six different features for a game simultaneously without causing a single collision.

Background agents persist throughout long sessions to continually build context, while every action gets logged locally to ensure seamless crash recovery. In stress tests running on NVIDIA Hopper hardware, the setup survived over 1,000 tool calls across a 24-hour run.

Not the smartest model, but dangerously cheap

Meta's self-reported benchmarks claimed Muse Spark 1.2 nearly matched top-tier models like Opus 5. However, independent tech reviews quickly pointed out that Meta had compared it against a trimmed-down GPT 5.6 Terra rather than the flagship Soul model.

On the independent Artificial Analysis leaderboard, Muse Spark 1.2 actually sits outside the top ten, trailing behind models like Quen 3.8 and GPT 5.6 Soul Max. It landed in a similar mid-tier position on Val's index for coding and financial tasks.

Yet, it has one massive selling point: price. Across task benchmarks, it ran significantly cheaper than Gemini 3.6 Flash and blew Opus pricing out of the water.

Meta AI chief Alexandr Wang told The Wall Street Journal that Muse Code offers an exceptionally strong choice for many workflows, specifically from a cost efficiency standpoint.

Why it matters

For months, Meta was seen as playing catch-up in the AI agent harness space against rivals like Anthropic with Claude Code and OpenAI with Codex. Muse Code proves that Meta is eager to control the developer interface, not just the underlying weights.

For developers, Muse Code likely will not replace top-of-the-line models for high-level architectural decisions. But if you have repetitive backlog tasks, repo-wide migrations, or test-suite generation, paying pennies for an agent that works in safe, parallel sandboxes makes a ton of sense.

We think Meta's approach here is remarkably pragmatic. Instead of waiting to magically beat the frontier models in pure IQ, Meta leaned hard into defensive engineering: isolated worktrees, local crash recovery, and dirt-cheap inference. It turns out that when an agent is affordable enough and cannot wreck your repository, plenty of engineering teams will happily let it run.