Meta launches Muse Code to take on Claude Code and Codex

Meta has released Muse Code, a new AI coding agent that can execute end-to-end programming tasks straight from the terminal. The tool is in beta for macOS and Linux and runs on Muse Spark 1.2, a new model tuned specifically for software development. With this, Meta is taking direct aim at rivals like Anthropic’s Claude Code and OpenAI’s Codex.

Muse Code isn’t a typical chatbot that spits out snippets. It can scan large repositories, draft a plan, modify files, and verify results. Meta is targeting long-running jobs where an AI needs to remember multiple steps and autonomously decide what to do next. The market for AI tools for developers just got a new contender.

The headline feature is persistent background agents. Many AI tools spin up a fresh sub-agent for each prompt, forcing a constant rebuild of context. In Muse Code, specialized sub-agents stay alive throughout the session. They can take follow-up actions on their own and decide when to report back to the main agent.

Meta says this cuts latency and duplicate work. The coder can also unleash multiple sub-agents on a single problem at once. That’s enticing for big projects—but it also raises the stakes for oversight: several fast agents can spread good fixes quickly, and bad ones even faster.

Another differentiator is a local event log. Every model call, tool action, approval, and change is recorded. If Muse Code or your machine crashes, the job can resume from the same point—no need to reconstruct what was already explored or edited.

Muse Code also ships with built-in commands. /plan creates a plan that the user must approve first. /grill hunts for weak spots in that plan, and /goal drives the agent toward a defined end target. These are especially useful for tasks that outlast a single Q&A round.

Under the hood, Muse Code runs on the new Muse Spark 1.2. Meta says it trained the model and the coding environment together. As a result, the model learns not just to generate code, but to use the agent’s tools, summaries, and sub-agents effectively.

According to Meta, version 1.2 is better at complex debugging, understanding full codebases, and handling projects with many sequential steps. In an internal test, the system executed over a thousand tool actions across up to 24 hours to optimize software for Nvidia Hopper chips. That primarily shows the agent can run for a long time; it’s not independent proof that Muse Code beats competitors on everyday projects.

Meta is going with pay-as-you-go pricing. The standard Muse Spark 1.2 rates are $1.25 per million input tokens, $0.15 per million cached tokens reused, and $4.25 per million output tokens. A token is a small chunk of text or code the model processes.

There’s also a cheaper contributor tier, where Meta may use user activity to improve its products. That might be fine for individual tinkering, but companies should be extremely cautious with source code containing sensitive logic, personal data, or security details.

A direct price comparison with Claude Code and Codex isn’t straightforward. Muse Code bills per token, while the other two also offer monthly subscriptions. Which is cheaper depends heavily on task volume, codebase size, and how many intermediate steps an agent needs.

For developers, Muse Code looks like an early, relatively low-cost challenger. The agent can be installed on macOS and Linux, and Muse Spark 1.2 is also available via Meta’s Model API. Windows support wasn’t mentioned at launch.

The beta label matters. Run the tool in a separate branch, limit permissions, and review every change before merging or deploying. An agent that can run tests still won’t catch every unintended side effect.

Even so, Meta just put a serious new player on the field. The mix of persistent agents, resumable tasks, and low usage pricing is distinctive enough to turn heads. Whether developers switch will come down to independent real-world tests, reliability on large repos, and the fine print around how Meta handles code and usage data.