Meta Muse Code

Quick Take

Muse Code is Meta Superintelligence Labs’ first coding agent, launched in beta on August 5, 2026. It’s a terminal-based tool, powered by Meta’s Muse Spark 1.2 model, built to plan, implement, and validate multi-file changes across large repositories, positioning Meta as a new entrant in a category already led by Claude Code, Codex, and Cursor.

Muse Code is a separate product from Muse Image, Meta’s earlier image-generation model, and from Muse Glimmer, Meta’s open-weight local model released a few days later. All three come out of Meta Superintelligence Labs and share the “Muse” naming, but they don’t do the same job: Muse Code is a closed, paid coding agent, Muse Image is an image generator, and Muse Glimmer is a downloadable, self-hostable open model you run yourself.

What It Does

Meta describes Muse Code as able to take on complete software engineering tasks across large repositories: planning a change, writing the code, and validating the result, rather than handling isolated single-file edits. It installs from the terminal with a single command on macOS and Linux.

Pricing: The Contributor Tier Is Worth Reading Closely

Muse Code’s pricing structure has an unusual feature worth calling out clearly. There are two tiers:

• Standard: $1.25 per million input tokens, $4.25 per million output tokens
• Contributor: $0.10 per million input tokens, $0.20 per million output tokens, roughly a 20x discount, in exchange for allowing Meta to train future models on your prompts and completions

That’s a real tradeoff, not just a promotional discount. If you or your organization work with proprietary, confidential, or client code, the Contributor tier’s data-training terms deserve a careful read before opting in, regardless of how attractive the price difference looks. The Standard tier keeps that data out of training, at the higher, more typical per-token price.

Installation and Platform Support

Meta documents installation via a one-command shell script on macOS and Linux. As with any pipe-to-shell installer, it’s worth reading the script and the current authentication terms on Meta’s beta page before running it. Windows isn’t officially supported; running it under WSL2 has been reported by some users but isn’t documented or guaranteed by Meta.

How It Fits the Coding-Agent Landscape

Muse Code enters a category with established players, each with a different angle:

• Claude Code, Anthropic’s terminal-based coding agent
• Codex, OpenAI’s coding agent, extendable with tools like Codex Router
• Cursor, an AI-native code editor
• Gemini CLI, Google’s terminal coding agent
• OpenCode, an open-source, model-agnostic coding agent

Meta is the newest entrant here, so it has the least public track record. Whether Muse Code’s actual output quality holds up against these more established tools is something to test directly on your own codebase, not something to assume from the launch announcement.

Honest Limitations

• Beta status. Expect rough edges, changing pricing, and evolving capabilities.
• macOS and Linux only, officially. No documented Windows support at launch.
• Not open source. Access and pricing are entirely on Meta’s terms.
• The Contributor tier trades data for cost. A meaningful privacy consideration for any proprietary or client codebase, not just a simple discount.
• Newest, least-proven option in its category. Don’t assume it matches the maturity of tools that have been in production use longer.

Alternatives Worth Knowing

• Claude Code, a mature terminal coding agent with a longer track record
• Codex, OpenAI’s coding agent
• Cursor, an AI-native editor rather than a pure terminal agent
• Gemini CLI, Google’s terminal-based coding agent
• OpenCode, open-source and model-agnostic
• Muse Glimmer, Meta’s own open-weight, self-hostable alternative to a paid coding agent like this one

For how to think about routing coding tasks between different agents, see Match the AI Agent to the Risk.