Meta's Muse Code Agent Ships with Multi‑Agent Architecture & 4M Token/Min Limits

The AI coding arms race just added a new dimension. Meta’s Muse Code isn’t just a wrapper around a large language model—it orchestrates multiple persistent agents in isolated environments and logs every decision for recovery. For AI developers and researchers, this is a real‑world testbed for multi‑agent coordination at scale, backed by a model co‑trained explicitly for software engineering.

Meta has officially launched Muse Code, a terminal‑based AI coding agent now available in beta for macOS and Linux. Positioned as the company’s first major foray into agentic coding, the tool is built on Muse Spark 1.2, a new iteration of Meta’s model family that received significantly more training compute dedicated to coding tasks and was exposed to a broader range of development environments. The launch signals Meta’s intention to compete directly with established AI‑powered developer assistants like OpenAI’s Codex and Anthropic’s Claude Code.

“

The Standard tier charges $1.25 per million input tokens and $4.25 per million output tokens, with generous rate limits of 3,000 requests per minute and 4 million tokens per minute.

Muse Code’s architecture is designed to handle complex, multi‑file engineering challenges. After a single command‑line installation, the agent can plan changes, write code, and validate results while orchestrating multiple persistent sub‑agents that operate in parallel within isolated environments. This multi‑agent coordination is a differentiator, allowing developers to parallelize large refactoring jobs or cross‑repository migrations. An exhaustive local event log captures every model call, tool execution, approval decision, and edit, enabling sessions to resume seamlessly after a crash—a capability that speaks to enterprise‑grade reliability and auditability.

The pricing model is notably aggressive and split into two tiers. The Standard tier charges $1.25 per million input tokens and $4.25 per million output tokens, with generous rate limits of 3,000 requests per minute and 4 million tokens per minute. The Contributor tier, however, is where Meta’s data strategy becomes transparent: developers who consent to let Meta use their prompts and completions for future model training pay just $0.10 per million input tokens and $0.20 per million output tokens—a discount exceeding 95%. This tier is constrained to 60 requests per minute and 2.1 million tokens per minute, yet it directly values user‑generated coding data and is expected to attract individual developers, academic labs, and startups operating on tight budgets. The approach mirrors tactics seen in the consumer AI space (e.g., free tiers of large language models), but applied to a professional developer tool, it could rapidly amass a proprietary dataset of real‑world coding interactions that fuels the next generation of Muse models.

Market context is critical. The AI coding assistant market is already crowded: GitHub Copilot, powered by OpenAI’s Codex, has a vast user base; Anthropic’s Claude Code targets safety‑conscious teams; and startups like Replit and Codeium are racing to build specialized experiences. Meta’s entry leverages its infrastructure scale and the ability to co‑train models alongside the tool itself. Muse Spark 1.2 is also available through the Meta Model API with expanded global access, hinting that Meta may seek to monetize the model independently, creating a platform play rather than just a product.