Muse Code
Meta's terminal coding agent — pay standard rate and it never trains on your code, or opt into Contributor pricing for a steep discount if you let it.
- Category
- Coding agents
- From
- $1.25/M input, $4.25/M output / per token
- Pricing model
- usage-based
- Free tier
- No
- Deployment
- Cloud
- Interface
- CLI
- Open source
- No
- Public API
- Yes
- Model / LLM
- Meta's own Muse Spark 1.2 (no br…
Facts re-verified
On this page
Our verdict
Muse Code is a fast-following terminal coding agent from a company with the compute and model budget to make Claude Code and Codex CLI defend their pricing, not yet a reason to switch daily drivers.
Its most novel mechanic is priced rather than toggled: a Contributor tier bills roughly 12x cheaper on input and 21x cheaper on output in exchange for letting Meta train on your prompts and completions, a real price-for-data trade Claude Code and Codex CLI only ship as a free opt-out setting.
The parallel sub-agent worktree model is a real capability edge for large multi-part jobs, but it ships as a beta, three months old, with no independently audited benchmark, no IDE or GUI surface, no spend cap of any kind, and rate limits that throttle the discounted Contributor tier 30x harder on requests than Standard.
Pick it to evaluate Meta's coding model at a real cost discount or to trial multi-worktree parallelism on a big refactor. Look elsewhere for an established daily driver with outside-verified accuracy and a hard ceiling on the bill.
What is Muse Code?
Muse Code is Meta's own terminal coding agent, in public beta since August 5, 2026. It plans a change, writes the code, and validates the result across large repositories rather than single-file edits. It installs from the terminal with one command on macOS or Linux.
Windows developers need WSL2. It runs on Muse Spark 1.2, a coding-focused update to Meta Superintelligence Labs' own multimodal reasoning model, reachable through Meta Model API or OpenRouter rather than as a bring-your-own-model harness.
Its most distinctive mechanic is priced, not just toggled: a Contributor tier bills roughly 12x cheaper on input and 21x cheaper on output than Standard in exchange for letting Meta train future models on your prompts and completions.
What does Muse Code do?
You install Muse Code with a single terminal command and point it at a repository. It plans a change before writing any code, then, for large jobs, fans the work out to sub-agents that run simultaneously in isolated git worktrees so multiple parts of a change can proceed without conflicting. Reviewer and implementation agents can keep working in the background while you continue in the terminal.
A local event log tracks every model call, tool use, approval and edit so an interrupted or crashed session can resume from where it left off. Three observer agents also run by default, each making its own billed model calls. Every request bills through Meta Model API or OpenRouter at either the Standard (no-training) or Contributor (discounted, trains on your data) per-token rate, and there is no spend cap.
How Muse Code works
- Install with a single terminal command (curl -fsSL https://dev.meta.ai/install.sh | bash) on macOS or Linux. Windows requires WSL2, there is no native build.
- Authenticate in the browser, then point Muse Code at a repository and describe the engineering job.
- For large jobs, it fans out to sub-agents working in parallel inside isolated git worktrees. The fan-out cap scales with the machine's core count, clamped between 2 and 16.
- Run /plan to generate an approval-gated plan, /grill to stress-test it, and /goal to keep the session aimed at the objective. Implementation and review continue in the background while you keep working.
- A local event log records every model call, tool use, approval and edit so a session can resume after a crash. Three observer agents run by default and add their own billed model calls.
- Every request bills through Meta Model API or OpenRouter at either the Standard (no-training) or Contributor (discounted, trains on your data) token rate. There is no spend cap.
Key features
- Parallel sub-agent worktrees
- Fans a large job out to multiple sub-agents that work simultaneously in isolated git worktrees; fan-out cap scales with machine core count minus two, clamped between 2 and 16.
- Persistent background agents
- Reviewer and implementation agents keep working in the background while you continue in the terminal, rather than blocking on one linear turn at a time. Pressing Esc stops the current turn only and does not halt background terminals or sub-agents.
- Resumable event log
- A local log records every model call, tool use, approval and edit, so a crashed or interrupted session can resume from where it left off rather than restarting.
- Priced, opt-in training tradeoff
- A Contributor tier prices tokens roughly 12x cheaper on input and 21x cheaper on output than Standard, with cached input another 75x cheaper, in exchange for letting Meta use your prompts and completions to improve future models. Standard carries a stated no-training commitment.
- Built-in /plan, /grill and /goal commands
- Three named commands rather than a single chat loop: /plan produces an approval-gated plan before any code, /grill stress-tests the plan, /goal drives the session toward the stated objective.
- Reach through Meta Model API or OpenRouter
- Muse Spark 1.2 is reachable from Muse Code, Meta Model API and OpenRouter, with browser-based authentication rather than a separate API key setup.
What are Muse Code's use cases?
- Large multi-part refactor
- A developer asks Muse Code to update several related features at once; it fans the job out to sub-agents in separate git worktrees so each part proceeds without stepping on the others.
- Background review while you keep working
- A reviewer agent watches an in-progress implementation in the background, so a developer can keep working in the terminal instead of waiting for a review pass to finish.
- Budget-conscious high-volume usage
- A team opts into Contributor-tier pricing to cut token costs roughly 12x on input and 21x on output for a high-volume workload where the code involved is not sensitive, accepting that Meta will train on those sessions.
- Long-horizon kernel or compiler tuning
- Meta's own benchmark iterated over 1,000+ tool calls across 24 hours to optimise a GPU kernel on NVIDIA Hopper hardware, the kind of multi-hour task the resumable event log and persistent background agents are designed for.
Who is Muse Code for?
- Developers who want to evaluate Meta's own coding model without switching their whole toolchain
- Teams comfortable running a days-old beta CLI who value large-repo, parallel-worktree execution
- Cost-sensitive users willing to trade code privacy for Contributor tier's steep per-token discount
Not forWindows-native developers with no WSL2 setup. Teams that need an established independently benchmarked track record rather than a three-month-old beta. Buyers who need a hard spend cap or a flat subscription plan. Teams that want a dedicated IDE or GUI surface rather than terminal-only access.
What does Muse Code integrate with?
- Meta Model API
- OpenRouter
- Web search grounding billed at $2.50 per 1,000 queries
Why use Muse Code?
- The training-data tradeoff is priced, not just toggled: Contributor tier trains on your prompts and completions and bills roughly 12x cheaper on input and 21x cheaper on output than every rival terminal agent, which only ships a free settings toggle. Switching to the Standard tier turns training off at full price, a real quantified choice a buyer can make deliberately.
- Parallel sub-agent execution in isolated git worktrees lets multiple parts of a large job proceed without merge conflicts, with a fan-out cap that scales with the machine's core count.
- Persistent background agents plus a resumable local event log mean a crashed session does not lose its place. Three named commands (/plan, /grill, /goal) structure planning, stress-testing and goal tracking.
- Reachable through Meta Model API and OpenRouter, with browser-based authentication rather than a separate API key setup.
- Backed by Meta's own compute and model budget and positioned explicitly on cost versus Claude Code and Codex CLI.
What are Muse Code's pros and cons?
What's great
- The training-data tradeoff is priced, not hidden in a settings toggle: switching from the discounted Contributor tier to the no-training Standard tier is a real, quantified choice at a stated cost.
- Parallel sub-agent execution in isolated git worktrees lets multiple parts of a large job proceed without merge conflicts, with a fan-out cap that scales with the machine's core count.
- Persistent background agents plus a resumable local event log so a crashed session does not lose its place, plus three named commands (/plan, /grill, /goal) for planning, stress-testing and goal tracking.
- Muse Spark 1.2 is reachable through Muse Code, Meta Model API and OpenRouter with browser-based authentication, which gives a buyer a second route if Meta's own portal is unavailable.
- Backed by Meta's own compute and model budget and positioned explicitly on cost versus Claude Code and Codex CLI.
Watch-outs
- No spend cap, no hard stop, no budget limit. The only brake is an email alert, and the payment threshold rises as you demonstrate you can pay.
- Contributor tier is throttled to 100 RPM against Standard's 3,000 — a 30x tighter request cap. An agent that fires many parallel tool calls hits that ceiling before a token ceiling.
- No native Windows support — Windows developers need WSL2 to run it at all.
- Terminal-only, no dedicated IDE plugin or GUI surface, unlike Cursor, GitHub Copilot or Windsurf.
- Locked to Meta's own Muse Spark model with no bring-your-own-model option, unlike OpenAI Codex CLI's Ollama/LM Studio/Bedrock support.
- Contributor access can be revoked mid-project for "policy violations" without warning, per developer reports during the launch week.
- Three observer agents (memory, skill, goal) run by default and add their own billed model calls on top of the main session, switchable in settings.json under runtime_capabilities.
- No flat plan, no subscription, no seat pricing. Five of the six serious agentic coding CLIs ship a flat plan with a hard spend cap; Muse Code ships neither.
- Zero-data-retention is sales-gated and unpriced rather than a self-serve setting, which blocks security reviews that need a contractual ceiling.
- Reasoning effort defaults to xhigh (sixth of seven levels); reasoning tokens bill as output at $4.25/M, so the default is one of the most expensive settings on the dial.
Muse Code pricing
Purely usage-based, billed per token through Meta Model API or OpenRouter. No subscription, no flat plan, no spend cap. Standard: $1.25/$4.25 per 1M input/output tokens, $0.15/M cached, 3,000 RPM and 4M TPM. Contributor: $0.10/$0.20 per 1M input/output tokens, $0.002/M cached, 100 RPM and 3M TPM, in exchange for Meta using your prompts and completions to improve its models. Web search grounding is $2.50 per 1,000 queries on top. Zero-data-retention is sales-gated and not self-serve.
Standard
$1.25/M input, $4.25/M output / per token
- $0.15/M cached input
- 3,000 RPM, 4M TPM
- Prompts and completions are not used to train Meta's models
- Same product and feature set as Contributor tier, no seat or platform fee
- Reachable via Meta Model API or OpenRouter
Contributor
$0.10/M input, $0.20/M output / per token
- $0.002/M cached input — roughly 75x cheaper than Standard on cached input
- 100 RPM, 3M TPM — a 30x tighter request cap than Standard
- Meta uses your prompts and completions to improve its models
- Access can be revoked without notice for policy violations
- Same product and feature set as Standard — only the data use and rate limits differ
Web search grounding
$2.50 / per 1,000 queries
- Charged on top of tokens
- Flat rate regardless of context pulled per query
See current pricing on dev.meta.ai ↗Compare Muse Code alternatives →
Frequently asked questions
How much does Muse Code cost?
Does Muse Code train on your code?
Does Muse Code run on Windows?
What model powers Muse Code?
Is there a hard spend cap?
Is Muse Code open source?
Muse Code alternatives
- Claude Code
The established terminal-first agent with a multi-year track record and a real subscription tier (from $17/mo), versus Muse Code's days-old beta and pure per-token billing.
- OpenAI Codex CLI
A model-agnostic terminal agent that also supports local Ollama/LM Studio/Bedrock models, versus Muse Code's single-vendor lock to Meta's own Muse Spark.
- GitHub Copilot
An IDE-native assistant with a dedicated graphical surface and enterprise seat pricing, for buyers who want more than Muse Code's terminal-only interface.
- Compare the whole category
Every agent in Coding agents, side by side on the same fields.
Anything to add?
This page is researched and re-verified by our editors — but the people who actually use Muse Code know things we cannot see from the outside.
Report an issueSomething wrong, outdated or a dead link? Tell us and we re-verify.
Write a reviewUsed Muse Code? Your experience helps the next buyer decide.
Get a quote from Muse CodeTell us what you need and we make the introduction.
Get weekly Coding agents updates
New agents and fresh verdicts. One short email, every week. One-click unsubscribe.