Meta has released Muse Code, a terminal-based AI coding agent now in beta, alongside Muse Spark 1.2, a new coding-focused AI model. The launch puts Meta in direct competition with Anthropic's Claude Code, OpenAI's Codex and other agentic coding tools that have become central to how many developers build software.
Muse Code is designed to handle complete software engineering tasks across large code repositories: planning changes, writing code and validating results. It installs on macOS or Linux with a one-line command, though users must log in with a Meta account and provide billing details. CEO Mark Zuckerberg announced the release in a post on X.
The agent's main architectural feature is what Meta calls async background agents. Instead of spawning helper agents fresh for each task, Muse Code keeps specialized background agents alive for the entire session, avoiding redundant information gathering. For large jobs, it fans out to parallel sub-agents, each working in an isolated git worktree so the developer's own copy is never touched. Meta said it tested the agent building six features for a game simultaneously with no collisions.
Muse Code also maintains a local event log of every model call, tool run, approval and edit, making the runtime "replay-exact and restart-safe" -- if it crashes mid-task, it resumes from where it stopped without losing work. It ships with bundled skills such as /plan, which turns a task into an approval-gated plan, and /grill, which stress-tests the plan.
Muse Spark 1.2, the model behind the agent, is a coding-focused update to Muse Spark 1.1 with "significantly scaled up" training compute on coding tasks. Meta co-trained it with Muse Code itself, using the agent's trajectories to improve performance inside the tool. The company also used a self-improvement loop: Muse Spark 1.1 generated challenging coding environments, then graded candidate solutions to build a training dataset for the successor.
When the original Muse Spark debuted in April, it scored 77.4 on SWE-Bench Verified, trailing Claude Opus 4.6's 80.8 and Gemini 3.1 Pro's 80.6. As a demonstration, Meta pointed Muse Spark 1.2 at GPU kernel optimization and let it run for more than 1,000 tool calls over up to 24 hours on NVIDIA Hopper hardware. The agent wrote, compiled and profiled code, and Meta says it found "substantial improvements" over baseline implementations.
Muse Spark 1.2 is available through Meta's Model API in two pricing tiers. The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens, with prompts and completions not used for training. A "contributor" tier costs $0.10 per million input tokens and $0.20 per million output tokens -- roughly 12x and 21x cheaper -- but in exchange Meta may use the prompts and completions to train future models. Even the contributor tier requires a payment method on file, and it carries tighter rate limits: 60 requests per minute versus 3,000 on standard.
The launch is notable for what it does not mention: open source. Meta's Llama family of open-weight models had been downloaded roughly 1.2 billion times, but the new Muse models are proprietary, cloud-only and have no downloadable weights. Rivals OpenAI and Google have released their coding CLIs under open-source licenses, while Meta's tool is proprietary like Anthropic's.