Muse Code

Learning Objectives

• Understand what Muse Code does and how it differs from other terminal coding agents
• Identify its distinguishing features: async background agents and a replay-exact event log
• Read the Meta Model API pricing correctly, including the tier that trades your code for a discount
• Evaluate when Muse Code fits versus Claude Code, OpenAI Codex, or Gemini CLI

What Is Muse Code?

Muse Code is Meta's command-line AI coding agent, built for work that spans a large repository rather than a single file. Meta describes its scope as taking on complex software engineering tasks across large codebases: planning the changes, writing the code, and validating the result.

It arrived alongside Muse Spark 1.2, a coding-focused model that Meta co-trained specifically to work inside Muse Code. That pairing is the product's central design claim — the benchmark numbers Meta published describe the agent and model together, not the model on its own.

Core Capabilities

Async Background Agents

Muse Code's headline feature is specialized agents that stay alive for the whole session rather than spinning up and tearing down per request. On multi-step tasks this removes the repeated cold-start cost, which is where much of the waiting happens in agentic coding workflows.

A Replay-Exact Event Log

The runtime writes a local event log designed to be replay-exact and restart-safe. If the agent crashes partway through a long refactor, the session can be reconstructed exactly rather than restarted from scratch. This is a meaningful engineering choice: long-horizon agent runs fail often enough that recoverability matters more than raw speed.

Bundled Skills

Three commands ship with the agent:

• /plan — draft an approach before any code is written
• /grill — stress-test a proposed plan against edge cases
• /goal — set and hold a target across a long-running task

Long-Horizon Training

Meta says Muse Spark 1.2 was trained heavily on whole-repository generation and large end-to-end projects, and that it used the previous Muse Spark 1.1 to generate harder training environments for its successor — a self-improvement loop applied to training-data creation rather than to the model's own weights at run time.

Benchmark Performance

Pricing

The two tiers are the most interesting thing about the pricing. The Contributor tier is roughly 12-times cheaper on input and 21-times cheaper on output — an enormous discount by any standard. What you hand over for it is the right for Meta to train on the code you send.

For a personal side project that gap may be an easy trade. For work under a client contract, an employer's intellectual-property policy, or any license that restricts redistribution, it usually is not a decision an individual developer can make alone. Treat the tier choice as a legal question rather than a billing preference, and default to Standard unless you have explicit authority to do otherwise.

How It Compares

Muse Code's genuine differentiators are the persistent background agents and the replay-exact event log. Its genuine disadvantages are platform coverage and the absence of any free tier to evaluate it with — you are paying API tokens from the first run.

Strengths

• Session-persistent agents cut latency on long multi-step tasks
• Replay-exact event log makes a crashed long-horizon run recoverable rather than lost
• 1 million token context handles whole-repository work without aggressive chunking
• Co-trained model and agent is a coherent design rather than a generic model behind a wrapper
• Planning commands (/plan, /grill, /goal) push toward thinking before generating

Limitations and Considerations

• No Windows build, which excludes a large share of professional developers outside macOS-first shops
• Terminal only — no graphical interface and no IDE integration at launch
• Public beta, so behavior and pricing can change without notice
• No free tier to evaluate with; every run bills against the Meta Model API
• Benchmarks are self-reported on Meta's own harness, and place Muse Spark 1.2 second in every published comparison
• Closed weights, so nothing here can be self-hosted or audited
• The cheap pricing tier costs you your code, which is frequently not the developer's call to make

Key Takeaways

• Muse Code is Meta's terminal coding agent for large repositories, launched August 5, 2026 with the co-trained Muse Spark 1.2 model
• Its distinguishing engineering choices are session-persistent background agents and a replay-exact local event log for crash recovery
• Meta's published benchmarks are self-reported and rank its own model second to Claude Opus 5 in all three
• The Contributor pricing tier is roughly 12-times cheaper because Meta trains on the code you send; Standard does not
• Platform coverage is the practical blocker — macOS and Linux only, terminal only, and no free tier to trial it