Muse Code

Muse Code is Meta Superintelligence Labs’ answer to Claude Code: a terminal-only agent, released in beta on 5 August 2026, running on Muse Spark — a coding model co-trained with the harness rather than bolted to it afterwards.

It is a credible first entry, and it has one setting that can cost you your source code if you install it and start working. That section is below, and it’s the reason to read this page before running the install line.

Install link

  curl -fsSL https://dev.meta.ai/install.sh | bash
muse-code auth
  

The binary is statically linked: the installer needs only curl, mktemp and bash — no Node, no Python, no Homebrew. git isn’t needed to install but is required at runtime, because the worktree features are built on it. Auth is an OAuth browser flow, with a device-code URL if the browser doesn’t open.

As with any curl … | bash, you’re executing a remote script with your user’s privileges, sight unseen. Fetch it to a file, read it, then run it — the same two extra seconds you’d want from any vendor.

Check your billing tier before you point it at work code link

Muse Code is sold as plans (Everyday $5, High $15, Power $50) or pay-as-you-go through the Meta Model API. Two token rates matter:

That’s roughly 20× cheaper, and the price is your sessions — prompts, completions, and therefore whatever code you were working on. Reporting on the beta says the CLI lands on the Contributor tier after install, so opting out is something you do, not something that happens by default.

Whether or not that default holds in your install, the action is the same and takes a minute: confirm the tier before the first run against a private repository. For client work or anything under an NDA, the cheap tier isn’t a trade-off you’re allowed to make.

What’s actually new in it link

Persistent subagents. Rather than spawning a fresh helper per task and re-gathering the same context every time, subagents stay alive for the session, work in isolated git worktrees, and decide for themselves when to report back. The worktree isolation is the part worth stealing conceptually — parallel agents that can’t stomp on each other’s files.

Workflow orchestrates whole teams of those subagents against one large task, reviewing or building in parallel while you watch and steer from a single view.

Rewind — double-Esc — rolls the session back to an earlier point in the conversation. Cheap to use and the fastest fix when an agent has talked itself into a corner.

Underneath, a crash-safe event log keeps long-horizon repository work recoverable, and there’s inter-session messaging plus an SDK in developer preview.

Slash commands, with one gotcha link

/goal is a real slash command. Most of the things people call commands — plan, grill — are skills, invoked as /skill plan and /skill grill. The built-in skill ids are create-skill, doctor, git, grill, grill-and-record, import, manage-settings, plan, read-session and taste. If /plan returns nothing useful, that’s why.

Is it worth switching? link

Not yet, and not as a replacement. It’s a beta, the benchmark numbers are vendor-reported, and the ecosystem around Claude Code — skills, plugins, MCP servers — is years ahead. What Muse Code brings that’s genuinely its own is the worktree-isolated persistent subagent model and a single static binary with no runtime to install.

Worth an afternoon on a scratch repository, on the Standard tier. Not worth pointing at your employer’s monorepo this week.

Next link

Before pulling a local model, check what your machine can run → llmfit