Sign in
Muse Code logo

Muse Code

Beta

Meta's terminal-based AI coding agent with persistent async background agents

Visit websiteCompare
Conditional

Meta's first coding agent and a credible beta entry with genuinely novel architecture — persistent background agents, worktree-isolated sub-agent fan-out, and a crash-safe event log. Powered by the co-trained Muse Spark 1.2, it lands second only to Claude Opus 5 on Meta's own published benchmarks. Vendor-run numbers, beta status, and closed-source binary distribution temper the promise. Best for developers evaluating coding agents who want to test persistent-context architecture and aggressive contributor-tier pricing.

AI-Native

Agent loop + persistent background agents hold full coding-session context — the model reads, acts, and continues across steps without context reset.

Is it right for you?

Good for

  • Persistent multi-step coding sessions: background agents retain repo context across tasks, avoiding redundant exploration and reducing steering
  • Long-horizon autonomous tasks: event-log crash recovery enables 24-hour runs, and the agent resumes precisely where it stopped with no lost work
  • Parallel feature work: worktree-isolated sub-agents build multiple features simultaneously without collisions (Meta demoed six game features in parallel)
  • Cost-sensitive experimentation: contributor tier at ~$0.10 input / $0.20 output per 1M tokens undercuts every comparable coding agent
  • Multimodal repo work: accepts video fly-throughs and images as coding context, which no incumbent terminal agent ships as a first-class flow

Not good for

  • Production workflows requiring independently verified benchmarks: all published results are vendor-run in Meta's framework with no third-party reproduction
  • Windows parity: the Windows build lacks voice input and session messaging and uses PowerShell instead of Bash
  • Open-source or self-hosted pipelines: closed-source binary installed via shell script, no public repository, API-dependent, no weight access
  • Teams with strict data governance: contributor tier sends prompts and completions to Meta's training pipeline, and opting out costs 12.5x more on input

Our experience

We haven't tested Muse Code ourselves. This profile is based on public documentation, user reviews, and community feedback.

Pricing

Paid
Standard$1.25 in / $4.25 out per 1M tokens
  • No training on your data: prompts and completions are not used to train Meta models
  • 3,000 requests per minute / 4M tokens per minute per team
  • 1M token context window
  • Web search grounding: $2.50 per 1,000 queries
  • Cached input: $0.15 / 1M tokens
  • Same prices on Muse Spark 1.2 and 1.3
Contributor$0.10 in / $0.20 out per 1M tokens
  • Prompts and completions used to train future Meta models
  • 150 requests per minute / 3M tokens per minute per team
  • 12.5x cheaper on input, 21.25x cheaper on output vs standard tier
  • Cached input: $0.002 / 1M tokens
  • Aimed at prototyping and experimentation where training on your data is acceptable
View full pricing

No verdict changes yet

The clock starts day one. Changes land here as our verdict evolves.

Sources

Verification log

  • Content· Updated

    Automated agent

    Docs now say Muse Code runs on macOS, Linux and Windows from one codebase (stored: macOS and Linux only, no Windows build). Windows notes: PowerShell instead of Bash, UAC prompts during sandbox setup, no voice input or session messaging. Default model muse-spark-1.2, with 1.3 also available. Pricing page (dev.meta.ai/docs/pricing-rate-limits): Standard unchanged ($1.25/$4.25, cached $0.15, 3,000 RPM, 4M TPM); Contributor rate limit now 150 RPM (stored 100), 3M TPM, prices unchanged. Docs no longer use the word beta, but status left as beta pending an explicit statement. Description, strengths (dashes removed), audiences and tiers refreshed.

  • Pricing· No changes

    Automated agent

How we evaluate

Want this running in your business?

We tested this tool and we build with tools like it every day. Tell us the workflow and we will set it up, integrate it and hand it over working.

Tell us what you need (opens in new tab)See what we build

The conversation runs on Gnosari, one of the tools in this directory. A real conversation, not a sales script.