Muse Code

Product Introduction

1. Overview: Muse Code is a terminal-based AI coding agent developed by Meta Superintelligence Labs, built on the Muse Spark 1.2 large language model. It operates as a resident background agent for long-running development tasks.
2. Value: It automates complex software engineering workflows directly from the command line, enabling developers to delegate coding, debugging, and system optimization tasks that can run for hours without manual intervention.

Main Features

1. Resident Background Agent: Unlike typical chat-based AI, Muse Code maintains a persistent session context, allowing it to execute multi-step, long-duration tasks (reportedly up to 24 hours) without rebuilding context each turn.
2. Append-Only Local Event Log: All agent actions are recorded to a local, replay-exact log on the user's machine. This ensures crash safety and allows tasks to be resumed from the last completed action, providing robust auditability.
3. Multimodal 1M-Token Context: Leveraging the Muse Spark 1.2 model, it can process images, video files, and PDFs within a one-million-token context window, enabling tasks like generating code from visual mockups or video tours.

Problems Solved

1. Challenge: Automating lengthy, complex coding sessions that require persistent state and numerous tool calls, which is inefficient for traditional, stateless AI assistants.
2. Audience: Software engineers, DevOps specialists, and researchers working on system optimization, codebase refactoring, or generating applications from multimodal inputs.
3. Scenario: A developer can provide a video tour of a physical process and instruct Muse Code to generate a full-stack web application for managing it, with the agent handling hours of planning, coding, and debugging autonomously.

Unique Advantages

1. Vs Competitors: Unlike subscription-based AI coding tools (e.g., GitHub Copilot), Muse Code uses a pay-per-token pricing model (Contributor and Standard tiers), which can be more cost-effective for sporadic, high-volume tasks. Its terminal-native, agentic design contrasts with IDE plugin competitors.
2. Innovation: Its technical edge lies in the "replay-exact" local log architecture and the integration with Meta's cutting-edge, multimodal Muse Spark 1.2 model, designed for extended reasoning and execution over very long contexts.

Frequently Asked Questions (FAQ)

1. How much does Muse Code cost? Muse Code uses a pay-per-token model with two main tiers: Contributor (approx. $0.10/M input tokens, $0.20/M output tokens) and Standard (approx. $1.25/M input, $4.25/M output). There is no subscription or free tier.
2. How do I install Muse Code? Installation is a one-line command for macOS and Linux: curl -fsSL https://dev.meta.ai/install.sh | bash. This places the muse binary in ~/.local/bin. Windows requires WSL2 (Windows Subsystem for Linux).
3. What can Muse Code actually do? It functions as an autonomous coding agent that can plan and execute complex software tasks. Key skills include generating execution plans (/plan), critically examining its own code (/grill), and pursuing high-level objectives (/goa), all within long-running sessions.