Meta has released Muse Spark 1.2 and an early beta of Muse Code, its first purpose-built coding agent designed for long-running, multi-agent software development workflows.
Muse Code provides visibility into every subagent, tool call, intervention, and cancellation through a replayable event log, giving developers an auditable record of the agent’s activity.
Meta is also expanding global public-preview access to Muse Spark through Muse Code and the Meta Model API.
Muse Spark 1.2 is additionally available through OpenRouter.
The updated model is optimized for tasks such as multi-file refactoring, extended debugging sessions and software development projects that continue beyond a single prompt.
Meta said Muse Code was incorporated into the model’s training process, helping improve tool usage and plan execution. The model was trained across multiple coding-agent environments so it can also generalize beyond Muse Code.
Muse Spark 1.2 supports a one-million-token context window, allowing developers to keep dependency graphs, legacy code and thousands of files available within a single session.
The model also uses planning, goal conditioning, context compaction, and asynchronous and parallel tool calls to maintain direction during long-running tasks.
Meta is offering two access tiers. The muse-spark-1.2-contributor model is rate-limited by tokens over a rolling five-hour period rather than by request count and may be used to improve Meta’s products.
The standard muse-spark-1.2 model uses Meta Model API pricing of $0.15 per million cached input tokens, $1.25 per million input tokens and $4.25 per million output tokens.
Meta is also beginning to accept requests for zero-data-retention arrangements.
Muse Code can automatically break a larger project into separate tasks handled by parallel subagents.
Each write-capable subagent can operate inside its own isolated Git worktree, allowing multiple agents to modify code simultaneously without interfering with one another or altering the developer’s main working copy.
Each child agent’s activity is recorded in a local JSONL event log. Sessions can persist after a laptop closes or a process crashes, and the muse resume command can reconstruct the previous session from that log.
Muse Code also includes built-in skills for interface design review, structured questioning and project planning.
These tools include /taste, /grilling, /grill-with-docs and /plan, and Meta said they execute only when explicitly invoked by the developer.
