Luna

Transparency

Last updated September 24, 2026

This page exists to be honest about what Luna is, what it isn't, and how it works under the hood. No fine print. No weasel words. We wrote this because legal polish is easy, but real transparency means telling you what we don't know.

What Luna is

Luna is a free, open-source menstrual cycle tracking companion. You log your cycles by talking to an AI assistant — type things like "my period started today" or "I'm feeling crampy" and Luna records it. Over time, it learns your patterns and predicts when your next period might come.

The prediction engine starts with a broad population range, then learns your typical cycle and variability from usable logged history. It reports a likely window for the next cycle and stays deliberately cautious when history is sparse, variable, or possibly incomplete. Conditions widen or withhold estimates where the evidence supports caution; they do not create unsupported precision.

Luna is open source under the MIT license. You can read every line of code that handles your data and makes predictions. The full pipeline — database schema, prediction algorithm, AI tool definitions, API routes — is in one GitHub repository. Nothing is hidden.

What Luna is not

  • Not a medical device. Luna has no FDA clearance, no clinical validation, and is not intended to diagnose, treat, or prevent any medical condition.
  • Not contraception. Luna cannot tell you when you are fertile or when it is safe to have unprotected sex. The only FDA-cleared contraceptive app is Natural Cycles, which uses basal body temperature — Luna does not.
  • Not a healthcare provider. Luna does not replace a doctor, nurse, or therapist. If something in your cycle concerns you, talk to a qualified healthcare professional — not an app.
  • Not validated.We don't know how accurate Luna's predictions are. The automated suite and synthetic backtests verify the implementation, but the live dataset is too small for a responsible accuracy comparison. Predictions are statistical estimates and may be wrong. Automated regression tests check the implementation, but there is no clinical validation or publishable real-world benchmark.
  • Not HIPAA-compliant.While Neon (our database) offers HIPAA on their Scale plan, Luna doesn't use that plan. Supermemory claims HIPAA but provides no public BAA. HackClub has no HIPAA certifications. Health-adjacent data is stored on infra without healthcare-grade compliance guarantees.

How predictions actually work

Luna's current prediction engine is open source in forecast.ts. Here's exactly what it does:

1. Broad starting assumptions

With no personal history, Luna uses a deliberately broad population starting point drawn from prospective cycle studies. Published total variation is not treated as if it were certainty about one person. PCOS, thyroid conditions, endometriosis, and contraception do not receive invented means when the source only supports a directional association.

The numerical assumptions, source populations, and transfer limits are listed in papers/references.md. Several components are conservative engineering assumptions that still require prospective calibration.

2. Learning from your history

Luna estimates your typical cycle and your cycle-to-cycle variation. Personal observations gradually outweigh the starting point, but a variance floor prevents one or two similar cycles from producing a falsely precise prediction.

Only the 12 most recent usable intervals inform a forecast. The exact shrinkage strength remains an assumption to validate.

3. Uncertain logs

One very short or long interval is set aside as a possible duplicate, spotting entry, or missed log. It remains visible and is counted in the forecast basis. Repeated long intervals are treated as a personal pattern instead of being discarded.

Luna does not know whether a long gap is a missed log or a genuine long cycle. This is an uncertainty rule, not a diagnosis.

4. Prediction intervals

Luna gives an 80% likely window for the next observation, not a confidence interval around an average. The range includes both estimated personal variability and uncertainty about the person's typical cycle. Ovulation is clearly labelled as a calendar estimate or withheld when unsuitable.

Synthetic tests check calibration under known assumptions. The live dataset is too small to establish real-world coverage.

How we process your data

Your data touches three third-party services. Here is the unvarnished truth about each:

Neon (database)

All structured data — cycles, accounts, chat messages, predictions, AI traces — lives here. Neon is a serverless PostgreSQL platform on AWS (8 regions, 4 continents).

Encryption at restAES-256In transitTLS 1.2+CertificationsSOC 2, ISO 27001HIPAAScale plan onlyParent companyDatabricks (acquired 2025)

Neon does not sell personal data. Core engine is Apache 2.0 open source. Luna uses the HTTP serverless driver with cache: no-store. Trust Center · Security

HackClub (AI proxy + search)

US 501(c)(3) nonprofit. Provides free AI and search to its community. Luna's AI proxy sends prompts to OpenRouter, which may route them to xAI or Anthropic. Web search requests use HackClub's search API.

Critical concern: HackClub logs every AI prompt and response in full (jsonb), linked to user ID and IP. No retention policy. No deletion schedule. No service-specific privacy notice. This means health information in your chat messages is stored indefinitely on their infrastructure.

All code is open source — the logging is verifiable. But it is not optional. If you need strong privacy, self-host Luna with your own AI backend. HackClub Privacy & Terms

Supermemory (personal facts)

When you say "remember I have PCOS," that fact goes to Supermemory v4 API. Luna sends only personal facts — never cycle data or chat messages. Scoped per user via containerTag.

InfrastructureTimescale + CloudflareAt-rest encryptionNot documentedComplianceClaims only (no audits)Third-party AIMay use OpenAI/Gemini

Supermemory Inc. is early-stage (founded by Dhravya Shah). Core engine is MIT open source. Luna uses a 3-second AbortSignal timeout on all Supermemory calls. Supermemory Privacy

Data flow summary

Data typeStored inAlso processed byLogging
Cycle recordsNeon (AWS)—SOC 2 audited; no HIPAA
Chat messagesNeon (AWS)HackClub → OpenRouter → xAI/AnthropicFull prompt+response logged by HackClub
Personal factsSupermemoryPossibly OpenAI/GeminiNo at-rest encryption documented
Web searches—HackClub search APISubject to HackClub's handling
Auth credentialsNeon (AWS)—bcryptjs hashed

Known limitations

  • No publishable real-world benchmark. Luna now has a leakage-free evaluation harness and synthetic benchmarks, but only four live users have retrospective targets. That is below our minimum reporting threshold, so real-data accuracy stays unknown.
  • AI can misunderstand you.Chat-based logging is convenient but imperfect. The date normalizer supports natural language ("yesterday," "May 3"), but the AI might misinterpret what you say or miss information.
  • Population evidence is not personal truth.The starting assumptions come from selected research populations and may not transfer to you. The model's variance split and shrinkage strength remain engineering assumptions to calibrate.
  • Condition evidence is limited. Luna no longer assigns unsupported means to PCOS, PCOD, endometriosis, thyroid conditions, or every hormonal method. It usually widens or withholds an estimate, but those choices still need evaluation.
  • Distribution assumptions. Cycle length uses a log-normal predictive distribution and bleeding duration uses a bounded normal approximation. Real personal patterns can still be multimodal or change over time.
  • Tests are not clinical validation. Eighty unit and regression tests cover the engine, validation, and leakage invariants. Authenticated browser journeys and user studies remain incomplete.
  • Limited with very little data. With 1-2 cycles, predictions are mostly driven by population averages — which may not match you at all.
  • No wearable integration.Luna cannot read data from Oura Ring, Apple Watch, or any wearable. If you track basal body temperature, you'll need to enter it manually.
  • Uncertain-log rules are heuristic. One extreme gap is set aside as possibly incomplete; repeated long gaps become a personal pattern. The app still cannot know whether a particular gap is a missed log or a genuine long cycle.
  • HackClub logs everything. Every message you type to the AI — which may include symptoms, cycle details, health questions — is stored indefinitely, linked to your identity, with no documented deletion schedule.

These limitations are documented in detail in Luna's papers directory, particularly the nontechnical paper and technical paper's Section 8 (Limitations), Section 6 (Experiments and evaluation), and Section 7 (Results).

What we don't do with your data

  • Sell your data. To anyone. Period.
  • Show you ads. No ad networks. No ad tracking.
  • Share data with insurers, employers, or brokers.
  • Train AI models on your conversations. The model providers (xAI, Anthropic) may have their own policies.
  • Use tracking cookies or analytics. Only essential cookies (NextAuth.js session JWT) and localStorage for preferences.
  • Share cycle data with a partner or doctor.
  • Connect to wearables. No Apple Watch, Oura Ring, etc.

Data retention & your rights

  • Images auto-deleted after 7 days. Cron job purges expired uploads from PostgreSQL.
  • Export anytime. Download all cycle data as Luna JSON v1 (mStart, mEnd, ovulationDate, cycleLength, periodLength, notes).
  • Account deletion. Coming soon to settings. Currently by contacting us.
  • Third-party logs.HackClub retains AI prompts and responses indefinitely — outside Luna's control.

Open source

Luna is MIT-licensed at github.com/a3ro-dev/luna. You can:

Read exactly how predictions are calculated
See every database query and API route
Inspect how data is sent to each third-party service
Review every forecast assumption and its evidence
Run your own copy with your own infrastructure
Read the 67 technical paper pages in papers/

Same AI model for all plans

Luna has three pricing tiers: Free, Premium at $5/month, and Premium+ at $12/month. All use the same Grok 4.3 chat model through the HackClub proxy, the same cycle data, the same forecast service, and the same chat tools. Plans change page composition, color, navigation, and conversational tone. Free is calendar-first; Premium adds guided navigation and a chat rail; Premium+ adds a split dashboard and wide-screen cycle context. The persona prompts are defined in models.ts.

Cookies & local storage

Luna uses only essential cookies (NextAuth.js session JWT for authentication) and localStorage (cookie consent choice, PWA install preference). No tracking cookies. No analytics cookies. No third-party cookies. No fingerprinting. The cookie banner at the bottom of the screen tells you this on first visit.

Contact

Questions, concerns, or want to discuss how Luna works?

akshatsingh14372@outlook.com

github.com/a3ro-dev/luna