Meta Muse Code: A New AI Coding Agent, Explained

Meta has released Muse Code, a terminal-based AI coding agent built to compete directly with Anthropic’s Claude Code, OpenAI’s Codex, and xAI’s Grok Build. Launched in beta on August 5, 2026 by Meta Superintelligence Labs, it’s Meta’s first serious entry into the fast-growing category of agentic coding tools — AI systems that don’t just autocomplete a line of code but plan, write, and validate changes across an entire repository on their own.

If you’ve read our beginner’s guide to AI agents, Muse Code is a concrete example of that shift applied to software engineering: instead of answering “how do I fix this bug?”, it goes and fixes it, then checks its own work.

What Muse Code Actually Does

Muse Code installs with a single shell command on macOS and Linux and runs from your terminal. Point it at a task — “add pagination to this API,” “fix the failing test in checkout.py” — and it plans the change, edits files, runs commands, and validates the result, iterating until the task is done or it needs your input.

It’s powered by Muse Spark 1.2, a coding-focused model Meta describes as trained on whole repositories rather than isolated code snippets, with a 1-million-token context window, context compaction for long sessions, and support for asynchronous, parallel tool calls — meaning it can kick off multiple actions (like running tests while editing a different file) without blocking on each one sequentially.

Parallel Subagents and a Crash-Safe Event Log

The standout feature is Muse Code’s ability to coordinate multiple persistent subagents on a single task, each working in its own isolated Git worktree so parallel changes don’t collide before they’re merged back. For a large refactor spanning dozens of files, that means the agent can split the work rather than grinding through it one file at a time.

Every model call, tool run, approval, and edit gets appended to a local event log that Meta calls “replay-exact and restart-safe” — if the process crashes or you cancel a run, you can resume from exactly where it left off rather than starting over. Muse Code also runs inside an OS-level sandbox with an approval workflow by default, so it can’t execute arbitrary commands without your sign-off.

Pricing: Two Tiers, One Catch

Muse Code’s pricing has a notable twist. The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens — broadly in line with premium coding-agent pricing elsewhere. But there’s a much cheaper “contributor” tier, at roughly $0.10 input / $0.20 output per million tokens, available if you opt in to letting Meta use your prompts and code to train its models.

That’s a meaningfully different trade-off than competitors typically offer, and it’s worth reading the terms carefully before choosing the discounted tier — especially for proprietary or client codebases.

How Does It Stack Up Against Claude Code and Codex?

Meta reports that Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1, a benchmark from the Laude Institute and Stanford researchers that tests agents against 89 real-world terminal tasks — things like system repair, data processing, and security fixes. By Meta’s own numbers, that puts Muse Code behind Claude Opus 5 (86.7%) but ahead of OpenAI’s Codex and xAI’s Grok Build on the same test.

Worth flagging: independent benchmark trackers have pointed out that there’s currently no third-party-verified leaderboard entry confirming Muse Code’s 82.9% score, and some analysts note that Meta’s published score for the previous model version (Muse Spark 1.1, claimed at 80.0%) came in noticeably higher than an independently reproduced score of around 76.2%. That doesn’t mean the new number is wrong — but it’s a reason to treat vendor-reported benchmarks as a starting point, not a final verdict, until independent labs confirm them.

Should Developers Try It?

If you’re already invested in Claude Code, Codex, or another agentic coding tool, Muse Code is worth a look mainly for two things: its parallel-subagent workflow for large, multi-file changes, and its aggressive contributor-tier pricing if you’re comfortable with Meta training on your code. It’s not positioned as the outright performance leader — by its own benchmark, it isn’t — but as a capable, competitively priced alternative in a market that’s rapidly filling out with options.

For teams evaluating agentic AI tools more broadly, it’s also a good case study in why the shift from generation to action is accelerating: every major AI lab now has (or is racing to ship) an agent that can act on a codebase autonomously, not just chat about it.

Takeaway

Muse Code marks Meta’s entry into a coding-agent race that already includes Anthropic, OpenAI, and xAI, betting on parallel subagents, crash-safe recovery, and unusually cheap pricing for developers willing to share their code for training. Its self-reported benchmark puts it just behind the current leader, Claude Opus 5 — but with no independent verification yet, that ranking is still Meta’s word for now. For developers picking an AI coding agent, the current field means it’s worth testing more than one on your own codebase rather than trusting any single benchmark chart.

---

Sources:

• TechCrunch — Meta launches Muse Code, an AI agent for large code bases
• 9to5Mac — Meta launches Muse Code AI coding agent for macOS and Linux
• MarkTechPost — Meta Superintelligence Labs Releases Muse Code (Beta)
• The Register — Meta wants to get inside your terminal with its new coding agent