Meta Launches Muse Code: A Terminal Coding Agent Backed by New Muse Spark 1.2 Model
Summary: Meta has entered the coding agent race alongside Anthropic and OpenAI, releasing Muse Code in beta — a terminal-based tool that plans, writes, and validates code across large repositories within a single persistent session.
Key Facts
- Muse Code: Free beta available on macOS and Linux via
curl -fsSL https://dev.meta.ai/install.sh | bash. It manages fleets of async background agents that persist for the entire session rather than spawning per task — reducing context loss on long development workflows. - Muse Spark 1.2: A coding-focused model co-trained with the Muse Code harness itself, featuring a 1-million-token context window and reported gains in code generation, complex debugging, and large-codebase navigation.
- Local audit log: Every model call, tool execution, approval, and edit is recorded in a local append-only event log, enabling replay-exact restarts after interruptions.
- API access: Muse Spark 1.2 is also available via the Meta Model API for developers and enterprise customers, with expanded global availability.
Why It Matters
Muse Code's arrival brings a credible third competitor into a market previously split between Anthropic's Claude Code and OpenAI's Codex. Meta's architectural bet on session-persistent agents — rather than per-task spawning — is a meaningful design choice for long-horizon engineering tasks. With major AI labs now all fielding terminal coding agents, 2026 is shaping up as the year this category becomes table stakes for AI platforms.
Further Reading
- CNBC report — CNBC
- VentureBeat technical breakdown — VentureBeat