Meta had the models; it now has the harness that binds them together. Its name: Muse Code.
Available in beta on Mac and Linux, Muse Code offers two modes of operation: an interactive terminal mode and a headless mode. Authentication can be done either through a browser or via an API key.
The default model is Muse Spark 1.2, released in tandem with the harness. Meta strongly advises not to switch it, and the rationale is simple: the large language model and the harness were trained together to work in concert.
Muse Spark 1.2 is described as a “moderate improvement” over Muse Spark 1.1, which had rolled out in early July and was initially only available in the United States. The optimizations focused on the coding layer, with particular attention to long-running tasks. As a result, capabilities such as planning, context compression, and adherence to stated objectives were honed and reinforced.
Paying with Your Data: an Option Not Available in the European Union
The introduction of Muse Code does not alter the pricing structure for Muse Spark models. The API maintains, beyond the standard tier, a so‑called “Contributor” tier. It offers a noticeable discount in exchange for two conditions. First, accepting stricter rate limits (60 requests per minute versus 3,000; 2.1 million tokens per minute versus 4 million). Second, granting Meta the right to use inputs and outputs to train future models. This latter point blocks the option within the European Union, and it is similarly unavailable in Australia, Brazil, Canada, South Korea, and the United Kingdom.
| (price per million tokens) | Input | Input (cached) | Output |
| Standard | $1.25 | $0.15 | $4.25 |
| Contributor | $0.10 | $0.002 | $0.20 |
Meta has also begun processing requests to activate the ZDR option (zero data retention).
Like Codex, No Nesting of Sub-Agents
Meta offers roughly a dozen cookbooks to illustrate Muse Code’s capabilities. These include natural-language planning, the distribution of tasks across sub‑agents, deterministic replays, objective tracking, and the immutability of guardrails.
The harness can manage up to 16 sub‑agents. The ceiling depends on machine resources (the general rule being the number of CPU cores minus two). As with Codex, these sub‑agents cannot spawn additional sub‑agents of their own. However, they can maintain their own Git tree to avoid write conflicts. Muse Code itself runs these sub‑agents to “observe” the execution. Each one operates along its own quality axis and can offer guidance to the principal agent without interrupting its flow.
By default, all shell commands execute within an OS sandbox, and a human must review those deemed risky (generally rm -f and rm -rf, with or without sudo). You can enable or disable these two aspects separately. There is a stricter mode (which requires review of any command not on an allowlist)… and there is a YOLO mode that also disables the sandbox and tells the harness to trust the workspace.
Muse Code includes a voice-input option (enabled by default on Mac; not on Linux). It allows manual management of context (compressing, summarizing the latest actions, exporting with the option to mask strings that resemble secrets…). The AGENTS.md files take precedence over CLAUDE.md files.
