Meta released Muse Code in beta on August 5–6, 2026 — the first terminal coding agent from Meta Superintelligence Labs (MSL), powered by the new Muse Spark 1.2 model. The install is one line:

curl -fsS https://dev.meta.ai/install.sh | bash

Muse Code is a fourth major closed-source terminal agent alongside Claude Code, OpenAI Codex, and Google Antigravity CLI. The architecture choices are what make it worth evaluating, even if the benchmarks don't lead the pack.

Sub-Agent Fan-Out Leaves Your Working Copy Untouched

Mark Zuckerberg's announcement thread describes the parallelization mechanism: "Muse Code runs specialized background agents that stay active your whole session, so they build up context over time instead of starting from scratch on every task. When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees. Your working copy is never touched."

Two mechanisms matter here:

  • Persistent background agents keep context across the session — previously inspected files, decisions made, patterns learned — instead of resetting the context window per task.
  • Git worktree fan-out spawns sub-agents each in their own git worktree, a primitive that checks out multiple branches to different directories simultaneously without touching the primary working copy. This is the same approach OpenCode's desktop client and several other harnesses use.

Zuckerberg's demo claim: Muse Code built six features for a game simultaneously with no collisions between sub-agents. Vendor-reported, not independently verified, but the mechanism is sound.

The Event Log Means No More Mid-Task Resets

The second architectural claim is arguably more consequential for production use: "Auditable by design. Every model call, tool run, and edit hits a local event log before it executes. If it crashes mid-task, it picks up exactly where it left off from that log. No lost work and no re-prompting."

This is a write-ahead-log pattern applied to agent actions. Anyone who has lost a long Claude Code /loop session to a crash knows the pain of re-explaining the task from scratch. Muse Code's event log replays state from the last logged action, not the last user prompt.

Meta also stressed a long-horizon stress test: 1,000+ tool calls over 24 hours on NVIDIA Hopper, running a kernel optimization task unattended. Again, vendor-reported, but the 24-hour autonomy claim is worth noting.

4.3 Points Behind Claude Opus 5 on Terminal-Bench 2.1

Meta published its own benchmark comparisons. The scores are self-selected, so treat them as a vendor's framing, not an independent leaderboard.

Benchmark Muse Spark 1.2 (Muse Code) Opus 5 (Claude Code, max)
Terminal-Bench 2.1 82.9% 86.7%
DeepSWE 1.1 59.3% 65.0%
Meta Internal Coding Bench 70.6% 79.4%

Muse Spark 1.2 trails Claude Opus 5 on every published chart, but comfortably beats Muse Spark 1.1 (76.2% on Terminal-Bench 2.1) and edges out GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%) on Terminal-Bench 2.1 specifically. The gap on DeepSWE is wider — 5.7 points behind Opus 5 and 5.5 behind GPT-5.6 Terra.

The Contributor Tier Trades Data for Cheap Access

Pricing details are secondhand — Meta hasn't published a pricing page yet. The story from public commentary: a contributor tier at roughly $0.10 per million input tokens versus about $1.25/M on the standard tier. In exchange, you allow Meta to use your usage data for model training. That's the same trade OpenAI and Anthropic offer on their discount tiers. Before pointing Muse Code at a proprietary codebase, confirm the exact terms and toggle settings at dev.meta.ai.

What I'd Test Before Switching

  • Open-source status unresolved. Zuckerberg's reply to the question: "I'll have more to share on that soon." Don't make procurement decisions based on an assumption either way. The model underneath (Muse Spark 1.2) is closed regardless.
  • Data training trade-off on the discount tier. If you're on the contributor tier, assume your code and prompts feed training data. Review the terms before running it on sensitive repos.
  • It's beta. The kernel-optimization and six-feature-parallel demos are vendor-reported, not third-party reproduced.
  • Benchmark gap is real. Expect to lose 4–10 points on typical coding benchmarks compared to Claude Code. Whether that matters depends on your task tolerance.

Muse Code is worth a weekend install if you want a free terminal agent with worktree isolation and crash recovery that you can run today. Just don't treat it as a drop-in replacement for Claude Code in production until you've validated it on your own repos.

Sources: