Meta Launches Muse Code Coding Agent, Bets on Spark 1.2 Model

Meta has joined the race for developers who write code with AI assistance. The company released Muse Code in beta, a terminal-based coding agent powered by the new Muse Spark 1.2 model, pricing access noticeably below rates charged by rivals Anthropic and OpenAI.

What Meta showed

Muse Code is Meta's first standalone coding agent, designed to work on large repositories. The tool plans changes, writes code, and validates the results, and installs with a single terminal command: curl -fsSL https://dev.meta.ai/install.sh, after which the user logs in via browser and runs the muse command in the project directory.

A key feature is the multi-agent architecture. When a task is large enough, Muse Code automatically splits it among subagents running in parallel, each given its own git worktree, which prevents code conflicts. In Meta's internal tests, the agent built six game features simultaneously without any collisions between changes.

Built-in commands

The agent offers four ready-made, explicitly invoked playbooks: /plan creates an action plan in the project's files and requires approval before execution, /grilling runs an interview that stress-tests design decisions, /grill-with-docs saves the outcome of that review into documentation, and /taste acts as a quality filter for the user interface. Meta stresses that none of these features trigger automatically - the agent won't reach for /grill just because a project looks weak.

All of the agent's actions, tool calls, approvals, and edits are written to a local event log in JSONL format. The log allows for auditing with jq and lets work resume after a session is interrupted, for example after a crash or a lost connection.

Model and pricing

Muse Spark 1.2 is a moderate update to July's Muse Spark 1.1, optimized for tasks typical of coding agents: multi-file refactors, long debugging sessions, and work that spans more than a single prompt. The model was developed together with the Muse Code harness, so it performs best there, though Meta says it also works in other coding tools developers already use.

Standard access through the Model API is priced at $1.25 per million input tokens and $4.25 per million output tokens, the same as Muse Spark 1.1. An alternative is the muse-spark-1.2-contributor identifier, billed in rolling five-hour token-limit windows at a lower rate in exchange for agreeing to let Meta use prompt and response content to further train its models. The company is also accepting requests for zero-data-retention access for enterprise customers.

Taking on Claude Code and Codex

The launch fits into Meta's broader developer-tools strategy: in April, the company replaced its Llama lineup with the Muse Spark model, and released version 1.1 in July. Muse Code currently has no dedicated graphical app, unlike Anthropic's Claude Code or OpenAI's Codex, and remains a purely terminal-based tool.

Releasing Muse Code in beta today. It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results - Mark Zuckerberg, CEO Meta

Meta's head of AI described the product's positioning bluntly as a play on price: this could be a very good option, especially from a cost standpoint. Muse Code's rates are close to or lower than those of its main rivals, and the cheaper contributor tier further lowers the barrier to entry for developers willing to hand over their data for model training.

For Polish development teams, this marks another low-cost alternative to Claude Code and Codex, particularly for large repositories where parallel subagents can shorten multi-stage refactoring work. The beta's current limitation to macOS and Linux, along with the lack of native integration with popular IDEs, could slow adoption outside teams already accustomed to working in the terminal.