Meta's Muse Code: a terminal agent built for repos most devs never touch

Meta shipped a coding agent for repositories that don’t fit in one developer’s head. Muse Code went into beta on Wednesday, runs in a terminal on macOS and Linux, and installs with a single curl command. Meta Superintelligence Labs built it against its own codebase.

Anthropic, OpenAI and Google got to the terminal first, so Meta arrives late with the cheapest meter. Muse Spark 1.2, the model underneath, bills $1.25 per million input tokens and $4.25 per million output, as Simon Willison logged on launch day. Meta’s pitch is scale plus price. Its own scorecards tell a more awkward story, because Muse Spark 1.2 finishes behind Claude Opus 5 on every coding benchmark Meta chose to publish, and the widest gap shows up on the bench Meta drew from its own pull requests.

What Muse Code actually does

Muse Code runs on Muse Spark 1.2, a coding-focused model Meta released the same day. Meta’s research post says the agent “takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results.” Background agents persist for a whole session rather than being spawned per task. A local event log appends every model call, tool run, approval and edit, which is what lets a killed run replay exactly and restart where it stopped.

Parallelism is the scaling trick. Mark Zuckerberg described how a big job gets carved up: “When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees. Your working copy is never touched.” He also told TechCrunch that a test run built six features for a game at once “with no collisions.”

There’s no GUI. 9to5Mac notes that unlike ChatGPT Codex and Claude Code, Muse Code ships no companion app, and Windows isn’t on the list. The pricing has a second tier worth reading twice: Willison flagged a muse-spark-1.2-contributor endpoint at $0.10 input and $0.20 output for developers who consent to data sharing, roughly a twelfth of the standard rate.

What the large-repo claim rests on

Meta’s evaluation methodology names the bench that carries the codebase argument. Meta Internal Coding Bench is 440 tasks taken from Meta’s own monorepo, derived from real internal pull requests covering bug fixes, feature development, refactoring and cleanup. Muse Spark 1.2 resolved 70.6% of them. Claude Opus 5 resolved 79.4%.

Public benchmarks repeat the pattern. On Terminal-Bench 2.1, the 89-task suite from Stanford and the Laude Institute, Decrypt tallied Meta’s chart at 82.9% for Muse Spark 1.2, 86.7% for Opus 5, 81.8% for Codex on GPT-5.6 Terra and 81.6% for Grok Build. DeepSWE 1.1, which spans 113 tasks across 91 repositories, put Muse Spark 1.2 at 59.3% against 65.0% for Opus 5. Meta’s launch post shows those results as bar charts without printing the numbers.

Meta attaches its own asterisk. The methodology document says: “We note that our evaluation setup (e.g., agent tools and system prompts) may not be specifically tuned for proprietary third-party models.” Independent numbers are thinner. Artificial Analysis, given pre-release access, measured 80% on Terminal-Bench v2.1 instead of 82.9%, and confirmed a 1M-token context window with cache hits billed at $0.15 per million. Nobody outside Meta has published a run against a repository the size of Meta’s.

What Meta hasn’t said

• Whether Muse Spark ever gets open weights. It has been closed since April, and Zuckerberg would only offer “I’ll have more to share on that soon,” per The Register.
• What the agent costs beyond tokens. No seat price, no included quota, no stated free tier.
• What the rate limits are. Meta’s post skips them entirely.
• Whether the large-repo claim holds for anyone else. Meta’s evidence for repo scale is a benchmark only Meta can run.

Skepticism landed fast. The Register picked up a commenter asking why Meta benchmarked against OpenAI’s mid-tier GPT-5.6 Terra rather than the higher-tier Sol. The question is fair. So is the trust question facing a company that spent the spring reorganizing its AI engineering ranks and now asks for write access to a working directory.

What this means for you

The reason to look at this is the invoice. If Claude Code or Codex already handles your work, Muse Code won’t out-score them, and Meta’s own charts say so. It might cost a fraction as much, which is the argument Meta AI chief Alexandr Wang made to the Wall Street Journal: “We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective,” in remarks TechCrunch relayed.

Two things deserve a real test before anything migrates. Run it on a branch of your actual repo, not a toy project, because whole-repository generation is the claim and nobody has verified it independently. Then kill the process mid-task and check whether the event log brings it back precisely where it stopped. That crash-safe replay is the one feature here that Claude Code and Codex don’t advertise, and it matters far more on a four-hour refactor than a three-point benchmark gap does. Skip the contributor tier on anything you don’t own outright.