treyleo16/ muse-code:latest

48 3 weeks ago

Muse Spark 1.2 for Coding.

cloud 
ollama run treyleo16/muse-code

Details

3 weeks ago

0242e0899ede · 30kB ·

system
You are Muse Code, an agentic coding CLI (command line interface) that helps users with software eng

Readme

Muse Code (Muse Spark 1.2)

Muse Code is Meta’s terminal-native agentic coding system and model environment engineered by Meta Superintelligence Labs. Powered by the backing Muse Spark 1.2 foundational reasoning model, Muse Code is designed for complex, long-horizon software engineering across multi-file codebases, combining terminal execution, parallel sub-agent fan-out, and continuous background validation.

Model & Agent Overview

Muse Code operates as a co-designed runtime agent and model architecture built specifically to compete with terminal coding environments like Claude Code and OpenAI Codex.

  • Developer: Meta Superintelligence Labs
  • Backing Model: Muse Spark 1.2 / Muse Spark 1.2 Contributor
  • System Type: Terminal Agent Runtime & Multi-Agent Orchestration Engine
  • Supported Platforms: macOS, Linux (CLI / Terminal)
  • Primary Use Cases: Full-repository refactoring, autonomous debugging loops, parallel feature development, and terminal/computer use automation.

Core Capabilities & Architectural Features

Parallel Sub-Agent Fan-Out & Isolated Worktrees

Muse Code handles multi-part software tasks by fanning out work across parallel, asynchronous child agents. To eliminate file collision and branch race conditions during multi-agent execution, each write-capable child agent operates within an isolated Git worktree.

Persistent Background Execution & Auditable Logs

Pairing a primary interaction loop with persistent, session-level background workers, Muse Code continuously reviews code quality and verifies changes without halting the developer’s terminal flow. All actions, git operations, and sub-agent execution trails are written to an auditable, restart-safe local event log.

Built-in Agent Workflows & Skills

Muse Code includes built-in commands tailored for agentic software planning: * /plan — Generates a structured, approval-gated architecture plan before executing writes. * /grill — Stress-tests proposed execution plans and architectural edge cases until logically sound. * /goal — Conditions sub-agents to work autonomously toward high-level end objectives.

Multimodal Vision-to-Code & Computer Use

Leveraging the multimodal vision capabilities of Muse Spark, the agent can capture UI screenshots, analyze visually rendered layouts or bug states, and map visual defects directly back to source code files for automated hotfixing.

Technical Specifications

Parameter Specification
Model String muse-spark-1.2 / muse-spark-1.2-contributor
Context Window 1,000,000 tokens (1M)
Input Modalities Text, Code, Vision (Images/Screenshots), Audio
Output Modalities Text, Code, Shell Execution Commands, Structured JSON
API Compatibility OpenAI SDK compatible endpoint via Meta Model API
Agent Features Git Worktree Isolation, Parallel Fan-Out, Context Compaction, Tool Grounding

Model Pricing Tiers

Meta offers two pricing tiers for the backing muse-spark-1.2 model engine:

Tier Input (per Mtok) Cached Input (per Mtok) Output (per Mtok) Data Usage
Standard (muse-spark-1.2) $1.25 $0.15 $4.25 Private (Prompts/outputs not used for training)
Contributor (muse-spark-1.2-contributor) $0.10 $0.002 $0.20 Discounted in exchange for model training telemetry