Muse Code – Meta’s Terminal Coding AI Agent Operational

Muse Code is a terminal coding AI agent developed by Meta and powered by the Muse Spark 1.2 model. It can handle complete software engineering tasks in large codebases, including planning changes, writing code, and verifying results. Muse Code’s background agent continuously accumulates context throughout a session, large tasks can be split into multiple sub-agents for parallel execution, and auditable local event logs enable recovery from the point of interruption after a crash. It directly competes with Claude Code and OpenAI Codex.

Full-stack engineering agent: Covers the complete software engineering workflow, including planning changes, writing code, and verifying results.
Persistent-context agent: The background agent remains active throughout a session, accumulating codebase context over time instead of repeatedly exploring the same areas.
Parallel sub-agents: Large tasks are automatically split into multiple sub-agents that run in parallel in independent worktrees without interfering with one another.
Working-copy protection: The agent runs in an isolated environment and does not directly modify the user’s original working copy.
Auditable logs: Every model call, tool execution, and edit is recorded in local event logs, ensuring full traceability.

Install the tool: Run the command curl -fsSL https://dev.meta.ai/install.sh | bash in a macOS or Linux terminal to complete the one-line installation.
Sign up and log in: Sign in to Muse Code with a Meta account and follow the prompts to add a payment method and activate the account.
Choose a plan: Select the pay-as-you-go plan or the Contributor plan, which costs as little as one-tenth of the standard price, based on your development needs.
Start a task: Enter a natural-language description of the software engineering task in the terminal. Muse Code will automatically plan changes, write code, and verify the results.

Long-running autonomous execution: Supports 24-hour uninterrupted kernel optimization tasks and continues to discover significant improvements after 1000+ tool calls.
Zero loss after crashes: Local logs enable checkpoint-based resumption, ensuring that no work progress is lost if a task crashes.
Low-cost entry: The Contributor plan costs approximately one-tenth of the pay-as-you-go price, and you can get started with just one line of code.
Enterprise-grade privacy: Supports zero-data-retention requests, and Meta promises not to use developer data to train models.

Project website: https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2

Large-codebase refactoring: Plan and execute cross-module code migrations in parallel without affecting the main branch.
Game feature development: Build multiple independent feature modules for a game project simultaneously, avoiding code conflicts and merge difficulties.
Kernel performance optimization: Run long-term automated tuning tasks to continuously identify opportunities to improve code performance.
Secure and compliant development: Use auditable logs and zero data retention to meet compliance requirements in sensitive industries such as finance and healthcare.
Full-stack automated delivery: Achieve end-to-end software engineering automation, from requirements planning to code writing and test verification.