How to Use Meta Muse Code: Setup, Workflow, and First Project Guide

Meta Muse Code is a high-performance terminal coding agent engineered by Meta Superintelligence Labs. Powered by the Muse Spark 1.2 foundation model, Muse Code is designed to operate directly inside your local developer terminal, executing end-to-end software engineering tasks: inspecting repository dependency graphs, writing cross-file patches, executing shell commands, running test suites, and autonomously healing build failures.

On standardized evaluations, Muse Spark 1.2 achieves 82.9% on Terminal-Bench 2.1, surpassing Claude Sonnet 5 (80.4%) by +3.1% relative and scoring 90.3% on MCP-Atlas tool execution.

This guide delivers a step-by-step walkthrough to install Muse Code, authenticate your developer credentials, initialize your first project, and configure multi-agent worktrees. Explore where Muse Spark ranks in our AI Model Leaderboard and read our Muse Code vs Claude Code Comparison.

Step 1: System Requirements & Installation

Muse Code runs natively on macOS (Apple Silicon & Intel) and modern Linux distributions (Ubuntu, Debian, Fedora, Arch). Windows developers should execute the setup inside WSL2.

1. Execute the Official Installer

Open your terminal and run the one-line shell installer:

curl -fsSL https://dev.meta.ai/install.sh | bash

[!NOTE] Security Best Practice: In production corporate environments, review the installer script before piping to bash: curl -fsSL https://dev.meta.ai/install.sh -o install.sh && less install.sh.

2. Verify Shell Path Configuration

Once the installer completes, reload your shell environment:

source ~/.bashrc  # Or source ~/.zshrc on macOS
muse --version

You should see confirmation of the active release: muse-code v1.2.0-beta (build 2026.08).

Step 2: Developer Authentication via dev.meta.ai

Muse Code authenticates through Meta’s developer API portal:

1. Navigate to dev.meta.ai and sign in with your developer credentials.
2. Under Model API Access, select Muse Spark 1.2 and generate an API key.
3. In your terminal, run the authentication command:

muse auth login

4. Paste your API token when prompted. The CLI securely encrypts and stores the key in ~/.config/muse/credentials.json with restricted file permissions (chmod 600).

Step 3: Project Workspace Initialization

Muse Code operates best when launched from the root directory of a clean Git repository.

Initializing an Existing Codebase:

cd ~/projects/my-production-app
git status  # Ensure working tree is clean or stashed
muse

Initializing a Greenfield Project:

mkdir auth-microservice && cd auth-microservice
git init
muse

Upon launch, Muse Code presents its interactive terminal cockpit:

┌─────────────────────────────────────────────────────────────┐
│  Muse Code v1.2 (Beta) • Engine: Muse Spark 1.2 (1M ctx)     │
│  Workspace: ~/projects/auth-microservice                    │
│  Git Branch: main • Status: Clean                           │
└─────────────────────────────────────────────────────────────┘
muse >

Step 4: Structuring Your First Agent Trajectory

The critical error developers make with autonomous coding agents is treating them like simple conversational chatbots. Muse Code is an execution loop: it plans, searches, modifies, and tests.

Phase 1: Repository Exploration (Read-Only)

Start by allowing the model to index the project without authorizing modifications:

muse > Inspect this repository architecture. Summarize the database schema,
identify the API route handlers, and report any failing test suites.
Do not modify any files.

Muse Code will scan your file tree, build an internal AST representation, and report back with a structured architecture summary.

Phase 2: Targeted Multi-File Feature Implementation

Once the model understands the project layout, assign a bounded engineering task:

muse > Implement a JWT token refresh endpoint in src/auth/refresh.ts.
1. Add the refresh route to the Express router in src/routes/api.ts.
2. Create unit tests in tests/auth.test.ts covering expired and invalid tokens.
3. Execute 'npm test' and verify that all tests pass.
4. Show me the git diff before finalizing.

Step 5: Advanced Features—Parallel Subagents & Worktrees

When handling large-scale refactors, Muse Code can fan out tasks across parallel subagents operating in isolated Git worktrees:

# Enable parallel subagent execution in config
muse config set agents.parallel 4

When instructed to refactor multiple modules, Muse Code creates separate worktree branches (muse/worker-1, muse/worker-2), executes modifications in parallel, runs tests across branches, and generates a merged patch. This prevents file collisions and dramatically accelerates multi-file migrations.

Crash Recovery via Append-Only Event Logs

If your terminal session disconnects or an out-of-memory error occurs 2 hours into a large task, Muse Code does not lose state. All prompt embeddings, tool calls, and AST diffs are logged locally in .muse/events.jsonl.

To resume an interrupted session:

muse resume --last

Muse Code parses the event log, reconstructs its reasoning graph, and continues from the exact step where execution paused.

Common Troubleshooting Scenarios

Final Takeaway & Best Practices

Meta Muse Code represents a major advancement in autonomous developer tooling. To maximize productivity:

1. Always start with read-only inspection before granting write permissions.
2. Require test verification in your prompt so the agent validates its own patches.
3. Review diffs carefully using muse diff before committing changes to main branches.

Compare Muse Spark against competing models in our AI Model Leaderboard and explore our comprehensive guide on the Best LLMs for Coding.