Which Muse Code model string is running?
Confirm whether traffic goes to Standard or the contributor model, then pin it in shared config. That one string separates a no-training statement from a training grant.
Muse Code privacy at a glance: Standard versus Contributor
Two Muse Code tiers, two data bargains, and a price gap that tracks the training grant.
| Tier | Data terms | Input / M | Output / M | Rate limit |
|---|---|---|---|---|
| Standard | Meta states it does not train on that tier's prompts and outputs | $1.25 | $4.25 | 3,000 RPM / 4M TPM |
Contributor (muse-spark-1.2-contributor) | You explicitly grant Meta rights to train on prompts and completions | $0.10 | $0.20 | 60 RPM / 2.1M TPM, some countries only |
Muse Code prices and tier terms are as reported, checked 6 August 2026; Meta's pricing and rate limits page renders client-side. Cost arithmetic stays on the Muse Code pricing page.
What the Muse Code Contributor terms actually say
Contributor pricing is not a discount with a data footnote attached: the data grant is what you pay with.
Selecting the muse-spark-1.2-contributor model string applies the terms: you permit Meta to use your prompts and completions to improve their products, and input drops to $0.10 per million tokens, output to $0.20. That data sharing is the consideration in the bargain, not incidental telemetry.
The open question: what "improve their products" covers
Training is the reading the tier name supports. Whether the wording also covers evaluation sets, red-teaming, abuse review or product analytics is unresolved, and Meta has not clarified that boundary anywhere we could find. The honest position on Muse Code privacy is that the scope is unclarified, not narrow. If your threat model separates "a reviewer may read this" from "a future model absorbed it", Muse Code terms do not either.
Muse Code data sharing travels with the request
Because the tier is a model string rather than an account setting, the choice is remade on every call: a copied script or an inherited environment variable can move a private repository onto the contributor path with nobody deciding to. Pin the Muse Code model string in shared config and watch the invoice — price is the tell.
Muse Code privacy on Standard: is the promise credible?
Distrust of Meta is the loudest theme in the launch discussion, and it deserves a hearing alongside the counter-case.
Many commenters in the Hacker News launch thread said they would not hand source code to Meta at any price, framing Muse Code privacy as a company-level trust judgement rather than a terms reading — a prior, not evidence about Muse Code itself.
The most cited counter-argument comes from Simon Willison in that discussion: a paid tier secretly training on customer data would be contract fraud, and a company with Meta's legal exposure has more to lose than it could gain. That is the strongest defence of Muse Code privacy on the paid tier.
A third reading sits between them, and it is the one to act on: "we do not train on this" is narrower than "we do not keep this". Retention windows, abuse logging and human review have no published answers. Muse Code privacy is therefore two claims at different confidence levels — a no-training statement, and an undocumented retention picture.
How the community frames it
One posted comment captures the Muse Code privacy sentiment this page answers with sources, not adjectives.
There's no way I'm giving Zuck any of my data
What the Muse Code event log keeps on your machine
Meta describes the Muse Code event log as local and append-only, and that design decision cuts both ways.
Meta's launch post for Muse Code and Muse Spark 1.2 describes an append-only event log that makes sessions replay-exact and restart-safe. One Muse Code privacy upside is real: session state lives on your disk, not a vendor dashboard.
The flip side is a full transcript of every run in your working environment — env files, keys pasted into prompts, customer data in fixtures — which backups and screen shares pick up. Meta has not published whether any part of a Muse Code log is synced: local by design, undocumented as to sync.
Muse Code privacy checks before you point it at a private repo
These apply to any hosted coding agent; Muse Code just makes two answers unusually explicit.
Does your employer allow hosted agents?
Muse Code sends whatever files the agent reads to Meta's API in order to work, so transmission comes before training in any Muse Code privacy review.
What retention answers can you get in writing?
Ask how long prompts are stored, whether human review exists, and how deletion is requested. If nobody answers in writing, that silence is your answer.
Can you test the blast radius first?
Run against a throwaway repository holding no secrets first. If an answer above is uncomfortable, Muse Code alternatives lists other trade-offs.
Muse Code privacy compared with Claude Code and Codex
Only Muse Code sells a documented discount in exchange for rights to train on your work.
| Agent | Documented data-for-discount tier | What to verify yourself |
|---|---|---|
| Muse Code (Meta) | Yes — Contributor grants training rights for roughly 12x/21x lower prices | Which model string your team runs; retention is undocumented |
| Claude Code (Anthropic) | None found in our review; subscriptions with enterprise controls | The data policy and admin controls on your plan |
| OpenAI Codex | None found in our review; subscription tiers plus a free tier | The data policy for your plan, free tiers included |
Competitor terms are deliberately not restated here: they sat outside our review and they change. Read each vendor's own policy. For features and price see Muse Code vs Claude Code and Muse Code vs Codex.
Muse Code account friction people report
Some barriers are identity and network problems, not data policy, but they shape the decision.
Muse Code sign-in runs through Meta's account system, which for some restates the objection: a work tool depending on a consumer social identity. Community reports describe a selfie verification step at sign-up — unverified by us, so treat it as community-reported. Meta's developer site did show a Login with Email option on 6 August 2026.
One further reported friction shapes Muse Code privacy reviews at work: corporate firewalls that block social domains can break authentication outright. Other sign-in failures live on our Muse Code troubleshooting page.
Related Muse Code privacy reading
Where to go next if Muse Code privacy terms opened a new question.
- Muse Code pricingThe cost gap between tiers.
- Muse Spark API accessWhich strings apply which terms.
- Our methodologyHow claims are verified and hedged.
Frequently Asked Questions
Does Muse Code train on your data?
It depends on the tier. On Standard, Meta states it does not train on that tier's prompts and outputs. On Contributor, you explicitly grant training rights on prompts and completions and get roughly 12x cheaper input, 21x cheaper output. The tier is set by model string — a per-request decision.
Is Muse Code privacy good enough for a private work repository?
That is a policy question your employer answers. Muse Code transmits whatever files the agent reads to Meta's API, as every hosted coding agent does. What is specific: Standard carries a no-training statement, Contributor does not, and retention and review policies are undocumented in the material we reviewed.
What does the Muse Code Contributor tier do with my prompts?
Contributor terms grant Meta rights to use prompts and completions to improve their products. Whether that covers only training, or also evaluation, red-teaming and analytics, has not been clarified. The discount is large enough to be selected accidentally in a script, so pin the model string if Muse Code privacy matters to your team.
Would you trust Meta with a coding agent?
The community is split. Sceptics treat Muse Code as a company-level trust call that no wording changes. Simon Willison's counter is that a paid tier secretly training on customer data would be contract fraud, an implausible risk at Meta's scale. The unresolved middle is retention, a narrower question than training.
Where can I read the Muse Code privacy terms in full?
Start with Meta's developer documentation, including the pricing and rate limits page where tier terms sit beside the prices. That page renders client-side, so some tools show it as empty. Everything here is reported as of 6 August 2026; anything unconfirmed is flagged, not smoothed over.
Does Muse Code data sharing explain why it is so cheap?
On Contributor, explicitly yes: the training grant is the stated consideration for the price. Standard pricing is not — it sits below several rival models per token while carrying the no-training statement. If a quoted Muse Code price looks impossible, check which model string produced it.
Can I withdraw data I already sent on the Contributor tier?
No withdrawal or deletion path is documented in the material we reviewed, and we will not invent one. Treat anything sent under Muse Code contributor terms as sent. If deletion rights matter, get that in writing first and keep sensitive repositories on the Standard model string.
Sources
- Meta AI Research — Introducing Muse Code and Muse Spark 1.2research.meta.ai
- Meta developer docs — Muse Spark pricing and rate limitsdev.meta.ai
- Hacker News — Muse Code launch discussion (278 points, 177 comments)news.ycombinator.com
- Meta for Developers — Muse Spark model pagedeveloper.meta.com