Meta Joins the Agentic Coding Race, and the Product Shape Is Already Decided

Meta just shipped its own agentic coding tool: software that writes, edits, and debugs code largely on its own, with a developer giving high-level direction rather than writing each line. It is the third frontier AI lab to launch a tool like this in roughly a year. The category has standardized on a product shape: terminal client, API key, pay-as-you-go tokens, no bundled IDE.

The new product is called Muse Code, and it runs on Meta's updated Muse Spark 1.2 model. Meta says the tool can run agents asynchronously in the background and execute complex coding tasks. Developers access it through Meta's developer website or through OpenRouter, the multi-model API gateway, with an API key and pay-as-you-go pricing. Pricing is $1.25 per million input tokens and $4.25 per million output tokens, unchanged from the previous Spark release.

That is the third product with this exact shape in about twelve months. Anthropic's Claude Code launched in 2025 as a terminal-native assistant that lives in the same workflow as a developer's existing editor. OpenAI's Codex rebranded and relaunched around a similar model: bring your own terminal, bring your own API key, pay per token. None of the three ships a full integrated development environment (IDE) the way Microsoft's GitHub Copilot or Anysphere's Cursor do. The setup is identical: agentic coding as a terminal-native, API-keyed, pay-per-token product.

A year ago, agentic coding was a demo, a research preview, or a hackathon curiosity. Today it is a default product category. Every major US frontier lab now has a product on the shelf that an engineer can sign up for in an afternoon. The competitive question has shifted from "can it run?" to "which model is best, and where is the distribution cheaper?".

There is a reason none of the three ships a full IDE. An IDE is a years-long incumbent business: keybindings, extensions, debugging integration, language servers, and team workflows. The frontier labs do not want to displace Cursor or Copilot. They want to be the model layer underneath both, and underneath whatever terminal a developer already has open. Selling tokens by the million is the cleaner business; competing with JetBrains is not.

The honest caveat: without independent benchmarks, there is no public basis to say Muse Code is better, worse, or equivalent to Claude Code or Codex on actual code quality. The CNET coverage is a clean re-report of Meta's announcement, not an evaluation. The structural read here holds on artifact alone: three labs, three terminals, three token meters, three API keys. Whether the underlying models are converging is a separate question that benchmarks, not launches, will answer.

For developers, Muse Code is a new vendor to test on Monday morning. The bigger signal: agentic coding is no longer a category question. The competition has moved to distribution.