Muse Code is Meta autonomous command-line coding agent designed to execute end-to-end software engineering workflows directly from the terminal. Built for multi-agentic development, the tool moves beyond simple inline completions by inspecting entire repository architectures, formulating explicit implementation strategies with user approval checkpoints, modifying source files across multiple packages, and verifying modifications via local test runners and linters.
The underlying intelligence comes from Meta's Muse Spark model family, which has a 1-million-token context window and accepts text, image, video, audio and PDF input; Muse Spark 1.3 is the latest hosted version, and Muse Code's default model is muse-spark-1.2, switchable with --model or /models. Meta also publishes open-weight models: Muse Glimmer, a 30B multimodal model distilled from Muse Spark for local and edge deployment under Apache 2.0, and Muse Spark Open (see below). The agent utilizes OS-level sandboxing and explicit workspace trust boundaries to protect local developer environments from unintended execution side effects.
Muse Spark Open and on-device use: on October 15, 2026 Meta released the weights of Muse Spark Open, an open-weight version of Muse Spark with 816B total parameters (42B active), a 1M-token context window and text, image, audio and video input, under the Muse Model License with limited commercial terms. Meta reports testing agentic workflows running entirely on-device using Muse Code powered by Muse Spark Open, including shipping iOS changes to the open-source Ice Cubes app and driving DaVinci Resolve. Local use needs serious hardware: Meta's run-locally recipe calls for two Mac Studios with M5 Ultra and 512 GB of unified memory each (or four with 256 GB), macOS 26.2 or later, MLX 0.30.1 or later and direct Thunderbolt 5 links, serving the roughly 537 GB NVFP4 checkpoint through mlx-lm as an OpenAI-compatible API. The full BF16 checkpoint is about 1.65 TB, and Meta's GPU serving recipes target 8x B300 (BF16) or 4x B300 (NVFP4). As of October 16, 2026 the Muse Code docs and changelog do not describe a dedicated local mode for connecting Muse Code to a self-hosted Muse Spark Open server; a standard install signs in to Meta Model API with a browser sign-in or API key. Muse Spark Open is also offered by hosted providers including Together AI, Fireworks AI, Baseten, Modal, Groq, Nebius and OpenRouter.
Muse Code features a multi-agent orchestration architecture that coordinates persistent background subagents to perform parallel tasks such as dependency analysis, test execution, and documentation lookup without blocking interactive developer flow. Built-in terminal commands include /plan for generating approval-gated change strategies, /grill for stress-testing and probing architectural plans, and /goal for driving complex objectives to completion. In addition, multi-session setups can exchange status and artifacts over Unix domain sockets.
Session reliability and transparency are maintained through an append-only local event log that records every model invocation, tool execution, and source patch, providing crash safety and replay-exact auditability. Muse Code operates on token-based pass-through billing via the Meta Model API on dev.meta.ai, offering both private Standard Tier and training-opted Contributor Tier access with zero upfront license fees.