Muse Code is Meta’s CLI coding agent, built by Meta Superintelligence Labs and launched in August 2026. It plans, edits files, and runs commands with an approval flow and an OS-level sandbox, and it is designed for both interactive terminal use and headless runs in scripts and CI.
models and pricing
Muse Code runs Meta’s Muse Spark model family, co-trained with the agent; other providers and local models are not supported. Subscriptions run from Everyday at $5/mo to Maximum at $200/mo, with usage multipliers at each tier, and pay-per-token billing through a Meta Model API key is also available.
what it’s good at
Muse Code fans large jobs out to persistent sub-agents in isolated git worktrees, and its headless mode suits scripts and CI pipelines. A TypeScript SDK and cross-session messaging support building on top of it, and it runs on macOS, Linux, and Windows.
Muse Code in crystl
Muse Code runs in a crystl shard with its own model selection and activity visibility, and a prompt can be handed to it at launch for one-shot worker runs. Crystl recognises the agent and tracks its activity, but the deeper surfaces are still unwired: its tool calls raise no approval cards, its sessions have no transcript crystl can read, and no resume command is documented for it.
A spawned Muse Code worker reports done when its process exits rather than on a turn-end hook.
The profile describes what crystl has tested; it never supplies a command, hook script, or launch setting. Those remain local to the Crystl installation.