Muse Code

Muse Code is Meta's command-line coding agent, released in beta on 5 August 2026 alongside Muse Spark 1.2, the coding model it was co-trained with. It puts Meta into the same category as Claude Code, OpenAI Codex and Gemini CLI: an agent that lives in the terminal, reads the repository it is pointed at, and takes on a whole engineering task rather than completing lines.

What distinguishes it

• Parallel subagents in isolated git worktrees. Fanned-out work goes to separate branches so subagents never collide on files, and your working copy stays clean. Enabled with --subagent-worktree-isolation; progress is tracked through subagent_status and subagent_wait.
• A replayable event log. Every subagent spawn, tool call, steer and cancel is recorded. Sessions persist as JSONL and muse resume picks one back up — the auditability argument Meta leads with, and a genuine difference from agents that forget what they did.
• Built-in playbooks as slash commands: /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds, /grill-with-docs records the reasoning, and /taste applies design defaults.
• A 1M-token context window on Muse Spark 1.2, aimed at whole-repository generation, multi-file refactors and long debugging sessions, with context compaction and asynchronous tool calls to keep sessions going.

Getting it running

Install is a single command on macOS or Linux — curl -fsSL https://dev.meta.ai/install.sh | bash — then authenticate in the browser. Approvals and an OS-level sandbox are on by default, so the agent asks before it acts outside its lane.

Cost

The standard muse-spark-1.2 model is $1.25 per million input tokens and $4.25 per million output, with cached input at $0.15. A contributor tier (muse-spark-1.2-contributor) is available in selected countries and is rate-limited by tokens in rolling five-hour windows rather than by request count — it is cheaper because your sessions may inform Meta's training, which is worth reading the terms on before pointing it at proprietary code. Zero data retention is available through Meta's sales team. The model is also reachable through the Meta Model API and OpenRouter.

Who it is for

Engineers working in large, long-lived codebases where a task spans many files and several hours, and teams that need to be able to reconstruct exactly what an agent did after the fact.