meta launches muse code, a terminal coding agent built to challenge claude code and codex

Meta released Muse Code in public beta on August 5, 2026. It is a terminal-based AI coding agent — not an IDE extension, not a chat interface, but a command-line tool that runs directly in your shell and is designed for large codebases and long-horizon engineering tasks. The model behind it is Meta's Muse Spark 1.2.

Available at launch for macOS and Linux. Windows support is not announced.

what it does differently

Muse Code's headline architecture decision is parallel sub-agents. Rather than a single agent working through a task sequentially, Muse Code spins multiple sub-agents simultaneously and coordinates their output. The positioning is explicitly about large codebases — the kind of codebase where a single-threaded agent spends most of its time reading before it writes anything.

The crash-recovery story is the other interesting piece. Muse Code logs every action the agent takes, which means if the agent (or your machine) goes down mid-session, it can resume from where it stopped rather than starting over. On long tasks — multi-file refactors, dependency upgrades, test generation sweeps — that matters. Losing progress on an hour-long agent session is one of the things that makes people distrust agents for serious work.

pricing

Standard tier: $1.25 per million input tokens, $4.25 per million output tokens. That puts it in the same range as mid-tier coding model pricing from Anthropic and OpenAI — competitive, not cheap.

There is a Contributor tier at $0.30 per million tokens total — roughly 21 times cheaper. The tradeoff is that Meta trains on the code you submit. That is a real tradeoff, not a footnote: if you are working on proprietary code, a codebase under NDA, or anything where the training data provenance matters to your employer or clients, Contributor tier is not a viable option. For personal projects and open-source work the math is different.

where this fits

Meta is framing this explicitly as competition for Claude Code and OpenAI's Codex — the two established terminal coding agents. That framing is accurate in terms of the category: all three are CLI tools that operate on your local codebase rather than living inside an IDE. The differences are in model quality, architecture, and the companies behind them, which carry different implications for data handling and long-term stability.

Claude Code has the strongest benchmark performance among terminal agents right now — Anthropic's SWE-bench numbers are well-documented. Codex has OpenAI's distribution advantages and enterprise relationships. Muse Code enters with a parallel-agent architecture claim and a pricing lever (the Contributor tier) that the others do not have. Whether the parallel sub-agent approach actually delivers faster or better results on real codebases is something that will need independent benchmarking.

the honest unknowns

This is a beta. Independent benchmark results are not yet available. The parallel sub-agent claim needs to be tested against the kinds of tasks that actually take a long time — not synthetic benchmarks, but the messy real-world work where agents currently struggle. Muse Spark 1.2 is also a new model without published evals, so the base capability is unverified against established coding benchmarks.

The Contributor tier's data terms need more scrutiny before anyone commits proprietary code to it. Meta's track record on data usage is complicated enough that this is worth reading carefully rather than accepting at face value.

Worth watching. Not ready to recommend over Claude Code or Codex for serious use until independent evals exist.