Muse Code
Meta's terminal coding agent that runs persistent parallel subagents across large repositories.
About
Meta Superintelligence Labs entered the terminal agent race on August 5, 2026 with Muse Code, a beta CLI built around the Muse Spark 1.2 model, which was trained alongside the tool itself. The agent plans, edits, and validates changes across large repositories by coordinating persistent asynchronous subagents that keep working in the background throughout a session, each in its own isolated working tree, with approval gates before risky actions land. Every model call, tool run, approval, and file edit is appended to a local event log, so a crashed or interrupted session can be restarted or replayed exactly. Bundled skills include /plan for task planning, /grill for stress-testing a change, and /goal for driving work toward an objective. A single curl script installs it on macOS and Linux. The model is closed and usage is metered: the standard tier costs 1.25 dollars per million input tokens and 4.25 dollars per million output, while a contributor tier drops to 0.10 and 0.20 in exchange for letting Meta train on prompts and completions.
Should you use Muse Code?
Pick it when
Pick it for long sessions on large repos where you want several background subagents working in isolated trees at once, and where replaying or resuming a crashed session exactly from the local event log matters more than choosing your model.
Look elsewhere when
Skip it on Windows, since the installer covers only macOS and Linux, or if you need a settled tool: it is a 2026 beta tied to one closed Meta model. The cheap contributor tier lets Meta train on your prompts, so private code fits better in OpenCode or Codex CLI.
Alternatives to Muse Code
- Claude Code
More mature terminal agent with subagents, hooks, plugins, and Windows support, but tied to a paid Anthropic plan instead of metered Meta tokens.
- OpenCode
MIT licensed and provider-agnostic, so you can swap models per task, where Muse Code locks you to Meta's closed Muse Spark model.
- Codex CLI
Apache-2.0 open source with sandboxed execution and Windows installers; runs on a ChatGPT plan or OpenAI key rather than Meta's metered pricing.
- Orca
Runs many existing CLI agents in parallel git worktrees from a desktop app, giving you parallelism without committing to a single beta agent.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- AI Code Assistants
- Price
- Paid
- Platform
- Local/Desktop
- Difficulty
- Easy (2/5)
- License
- Proprietary
- Added
- Aug 24, 2026