Muse Code defaults to the muse-spark-1.2 model. It supports interactive terminal sessions, CI pipelines, approvals, OS sandboxing, session messaging, and multi-agent orchestration via a workflow engine. A stable versioned session protocol is available via `muse serve`, with a TypeScript SDK (`@muse-code/sdk`) for programmatic access.

Muse Code now supports multi-agent workflows that can coordinate up to 1,000 child tasks (with up to 16 active in parallel) to run independent work streams simultaneously. A dedicated /workflows control room lets users monitor status, token usage, elapsed time, and child progress, with controls to pause, resume, skip, or restart individual child agents. Workflow availability depends on the workflow-script-engine-v8 build component and is currently unavailable on the aarch64-apple-darwin (Apple silicon macOS) artifact.

Muse Code now supports naming sessions and sending messages between live sessions for handoffs, review requests, and status updates. Each session maintains its own transcript, workspace, permissions, and task state, enabling coordinated multi-agent workflows. Sessions can also be isolated via Git worktrees when multiple agents need to write files independently.

Muse Code has published a dedicated Cookbook section covering advanced agent-building patterns including audit and resume, deterministic replay, staged approvals, sandboxing, immutable guardrails, subagent fanout, goals, bundled skills, scheduling, and side chats. This represents a significant expansion of the platform's orchestration and safety primitives for developers building production-grade agents. The content is framed as official harness building blocks for safety, recovery, and orchestration within the Muse Code ecosystem.

Muse Code now controls side effects through two independent layers: an approval/review system and an OS-enforced sandbox, both enabled by default. Shell commands are parsed and reviewed stage-by-stage rather than as a single line, blocking execution at the first dangerous stage. Scoped trust allows per-workspace trust decisions and prefix-based allow/deny rules, while workflow child agents inherit and can only narrow (never expand) the parent session's permission boundary.

Muse Code now offers a flat monthly subscription model as an alternative to pay-per-token billing, with 3 subscription tiers available. Users sign up via a Meta Account through the Meta Model API and Accounts Center, and the subscription is tied to a dedicated Muse Code API key. Additional API keys outside the subscription continue to be billed pay-as-you-go.