Meta has released Muse Code, a beta terminal-based coding agent designed to perform software engineering tasks across large repositories. The tool can plan changes, write code, and validate results.
Muse Code operates through a terminal rather than an application interface and can be accessed via a one-line command in a macOS terminal. The installation command is 'curl -fsSL https://dev.meta.ai/install.sh | bash'.
The agent uses a simple loop combined with asynchronous background agents, coordinating multiple persistent subagents for each task to resolve problems. It also includes default skills such as /plan, /grill, and /goal.
A local event log records every model call, tool run, approval, and edit, making the runtime restart-safe and replay-exact.
Muse Spark 1.2 integration
Muse Code is powered by Muse Spark 1.2, Meta's newest model and a coding-focused update to Muse Spark 1.1. Meta co-trained Muse Spark 1.2 with Muse Code.
To develop Muse Spark 1.2, Meta scaled up training compute on coding tasks and expanded training environment diversity. The training process included rejection sampled harness trajectories and recipe optimizations for goals, compaction, and subagents.
Meta also utilized Muse Spark 1.1 to generate instruction-following templates and coding environments to create a training dataset for the new model.
Muse Spark 1.2 was trained on long-horizon coding tasks, such as auto-research, large end-to-end projects, and whole-repository generation.
Performance and pricing
The model's ability to iteratively optimize GPU kernels was tested over 1,000 tool calls lasting up to 24 hours, with benchmarking performed on MLA and KDA kernels for NVIDIA Hopper GPUs.
Beyond code, Muse Code can interpret an mp4 fly-through video of a home to produce a booking and marketing page for vacation homes.
$1.25 / $4.25 per million tokens
Muse Code pay-as-you-go pricing for developers: $1.25 for input tokens and $4.25 for output tokens.
It maintains a specialized terminal-based workflow.
Updates
Meta CEO Mark Zuckerberg stated that Muse Code can plan changes, write code, and validate results across large repositories, even building six game features simultaneously without collisions by utilizing parallel sub-agents in isolated worktrees. Meta’s AI chief Alexandr Wang highlighted the tool's cost-effectiveness, as users can access it via a pay-as-you-go pricing model starting at $1.25 per million input tokens and $4.25 per million output tokens. The tool can be installed using a single curl command, though some concerns have been raised regarding its terminal-only interface and potential privacy implications.
Meta CEO Mark Zuckerberg stated that Muse Code can plan changes, write code, and validate results in large repositories by utilizing sub-agents that work in parallel in isolated worktrees without touching the user's working copy. The tool is available via a single command installation and follows a pay-as-you-go pricing model of $1.25 per million input tokens and $4.25 per million output tokens for contributor tier users. In performance benchmarks, Muse Spark 1.2 with Muse Code scored 82.9% on Terminal-Bench 2.1, outperforming GPT-5.6 Terra (81.8%) and Grok Build (81.6%), though it trailed Claude Code on Opus 5 (86.7%).
Meta CEO Mark Zuckerberg stated that Muse Code can perform complex planning and coding across large repositories by utilizing sub-agents in parallel isolated worktrees, even building six game features simultaneously without collisions. The tool is available via a single curl command and follows a pay-as-you-go pricing model of $1.25 per million input tokens and $4.25 per million output tokens on the contributor tier. Performance benchmarks show Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, outperforming GPT-5.6 Terra and Grok Build but trailing Claude Code on Opus 5, while it also recorded 59.3% on DeepSWE 1.1 and 70.6% on Meta's internal coding benchmark.
Meta CEO Mark Zuckerberg stated that Muse Code handles large repositories by fanning out to parallel sub-agents in isolated worktrees to ensure the working copy remains untouched, even successfully building six game features simultaneously during testing. The tool is available via a single terminal command and utilizes a pay-as-you-go pricing model for its contributor tier, costing $1.25 per million input tokens and $4.25 per million output tokens. While Muse Spark 1.2 achieved 82.9% on the Terminal-Bench 2.1—outperforming GPT-5.6 Terra (81.8%) but trailing Claude Code on Opus 5 (86.7%)—it also scored 59.3% on the DeepSWE 1.1 benchmark and 70.6% on Meta's internal coding benchmark.
Meta CEO Mark Zuckerberg stated that Muse Code can manage large repositories by fanning out to parallel sub-agents in isolated worktrees without touching the user's working copy, even successfully building six game features simultaneously during testing. The tool, which can be installed via a single command, follows a pay-as-you-go pricing model of $1.25 per million input tokens and $4.25 per million output tokens for users on the contributor tier. In performance benchmarks, Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, trailing Claude Code on Opus 5 (86.7%) but exceeding GPT-5.6 Terra on Codex (81.8%) and Grok Build (81.6%), while also scoring 59.3% on DeepSWE 1.1 and 70.6% on Meta's's
