Limbra is a mobile app
for multi-addiction consumption tracking grounded
in neuroscience. This dossier covers the vision,
the scientific specs, and the project's making-of.
Your feedback matters; pick the track that speaks to you.
The app is available on Google Play; feedback from real use is still very welcome.
Clinical feedback and research collaborations move this project forward faster than anything else.
The app's key screens, in half a minute; enough to get a feel for it before you try it.
Playing it loads the YouTube player ("no cookies" mode). Nothing is loaded from Google before the click.
The problem, the philosophy, the 5 pillars, the user journey, and the guardrails. Everything you need to understand why this project exists and how it works.
Over 115,000 deaths per year are linked to tobacco (75,000) and alcohol (41,000). The annual social cost of these two substances is estimated at €258 billion (€156bn for tobacco, €102bn for alcohol), plus €7.7bn for illegal drugs. In 2023, 246,000 hospitalizations were linked to an alcohol-related diagnosis.
Source: OFDT, Drogues et addictions, chiffres clés 2025 (France); social cost figures from Kopp, OFDT (2023), 2019 data.
The landscape of French-language apps is still dominated by single-substance tools and the day-counter logic. Limbra takes the problem differently: several kinds of use in a single app, three goals to choose from, no day counter, and data that stays on the phone.
Limbra starts from this observation to offer something radically different.
Addiction is not a lack of willpower. It's a brain learning process (Marc Lewis, 2015 and 2017); a debated model, which other researchers challenge in favor of a disease-based reading. Either way: not a moral weakness. Limbra takes a harm reduction approach.
The app does not diagnose and does not prescribe. In a crisis, it gives access to helpline numbers (3114, 15, alcohol helpline) and lets people save their own contacts, including a professional. Crisis tools are free and work offline.
Primarily people struggling with a substance or behavior, and the healthcare professionals who support them. To a lesser extent, the app also serves people with no addictive disorder who simply want to better understand and manage their consumption (social drinking, screens, etc.), via Observe mode.
An app available today on Google Play; content not final. What follows explicitly distinguishes what's already in the app from what's in the works.
A daily reading with four levels (starting up, calm, variable, storm), with an estimated sensitive time window when entries allow it. Today it runs on explicit rules applied to the person's entries: grouping by time, time of day, mood, and an automatic switch to "storm" during an active craving. A minimum of 4 days of data is required before any reading.
Multi-sensory pattern interrupt: haptic + visual + guided 4:6 breathing (vagal activation). The prefrontal cortex takes back control. A shortened demo is testable on the landing page.
A craving rises, then eases. For quitting smoking, the NHS says it usually lasts only a few minutes. The timer makes that duration visible, measurable, bearable. A 60-second post-craving replay to build metacognition.
3-to-5-minute micro-lessons: the reward circuit, dopamine (wanting ≠ liking), the prefrontal cortex, neuroplasticity. Sources: Koob, Berridge, Lewis, ACT.
No clean-days counter. The “I used” button is the same size as the others. Skills that stay with you (moments weathered, sensitive windows identified) instead of streaks.
Also available: Routines (small support habits, with reminders), the three-detail-level Journal, the Substance map, and the Profile (goal, trusted contacts, passcode lock, data export and erasure).
Weather currently runs on explicit rules, not on machine learning. A model trained on the person's own data (gradient boosting, heart-rate variability, sleep) still needs to be built. No performance has been measured to date: figures will be published after evaluation on real data, not before.
The app currently asks whether the person owns a connected wearable, but reads no data from it: the Health Connect integration is not built.
The educational "Transfers" module exists and explains the mechanism, but spotting one remains manual: the app runs no cross-analysis between different kinds of use.
The staged suggestions (triggers, contacts, professional follow-up), the monthly review, and the neuro mirror effect don't exist yet. The app currently handles startup with a simple threshold: no weather reading before 4 days of data.
Substances (alcohol, tobacco, cannabis, cocaine, opioids, MDMA, benzodiazepines, medication) + behavioral (screens, gambling, food, shopping, sex).
Not yet in the app. The educational "Transfers" module exists and explains the mechanism, but no cross-analysis between kinds of use is done: spotting one remains manual. The principle, once built: if substance A ↓ and behavior B ↑ over 14 days (≥30%), the AI would flag it gently: “Sometimes when you cut back on X, the brain looks for another source of stimulation through Y. That's normal; it's dopaminergic cross-tolerance. Want to talk about it?”
A planned journey, not yet built. What exists today in the app: onboarding, SOS available immediately, one-tap logging, the Learn modules, and a 4-day data threshold before the first weather reading. The staged suggestions, the monthly review, and the neuro mirror effect described below still need to be built.
Quick onboarding (7 screens, 45 sec) → first mood picker. The app doesn't overload you. SOS is available immediately.
Observation phase. Tap-to-log. The AI learns the patterns. Trigger suggestion on Day 3, contacts on Day 7. First neuro module.
Mini-weather over 7 days. Predictive model calibration, with or without a wearable.
Calibrated weather. Crisis kit prepared. Suggestion to seek professional follow-up if not already declared (Day 21). Monthly review on Day 30.
Refined predictions. Transfer detection active. Monthly neuro mirror effect. Acquired skills available to review. Goal adjustment possible.
Planned principles, not yet in the app. The app contains no AI assistant today; these safeguards are the ones the project commits to before building one.
The brain changes (neuroscience) AND can relearn (plasticity). No diagnosis, no label. A future assistant would never say "you're sick" or "dopamine is pleasure."
The emergency numbers and services below are for France only. If you are outside France, call your local emergency number or find a verified helpline in your country at findahelpline.com.
What the app does today. The app detects no danger signal: it does not analyze what the person writes and includes no AI assistant. Emergency numbers are one tap away from the SOS home screen, the lock screen and the four SOS phases; a button opens the phone dialer, and the person chooses whether to call. The “Être accompagné·e” (Getting support) screen also offers Drogues Info Service, Alcool Info Service and Tabac Info Service (France, drug, alcohol and tobacco helplines) and two personal contacts the person has saved.
| Situation | What the app offers today | Resource |
|---|---|---|
| Suicidal ideation | “Besoin d’aide immédiate” (Immediate help) screen: 3114 is offered first; 15 sits right below it on the same screen | 3114 (24/7, France) |
| Dangerous withdrawal (alcohol, benzodiazepines, opioids) | The 15 button on the emergency screen names difficult withdrawal. Guidance appears in the Learn fact sheets (see below). No alert is triggered automatically | 15 (SAMU, France) |
| Opioid + benzodiazepine + alcohol mix | Warning always shown on the emergency screen (potentially fatal mix, call 15), a naloxone reminder and a reminder to call 15 even after using it | 15 (SAMU, France) + naloxone |
| Minor in danger | 119 button on the emergency screen; 119 also appears on the “Violences” screen in the Profile | 119 (France, child protection line) |
| Violence, experienced by oneself or by someone close | Profile, “Aide & sécurité personnelle” (Help & personal safety), “Violences” screen: 17 or 112, 114 by text, 119, online reporting, 116 006, 3919 (France) | 17 / 112 (France) |
| Psychotic symptoms | Planned, not in the app: no specific referral today | CMP (planned; French mental health center) |
What the app says about withdrawal today. The Learn fact sheets on alcohol and on several benzodiazepines (diazepam, lorazepam, temazepam, flunitrazepam) state that stopping abruptly after prolonged use can be dangerous and should be planned with a doctor; the alcohol sheet adds to call 15 (France) without waiting in case of severe tremors, confusion or seizures. No message is triggered when a use is logged or when use stops: the planned alert before an abrupt stop is awaiting written clinical validation and is not reproduced here.
| Substance | Withdrawal severity |
|---|---|
| Alcohol (heavy use) | Delirium tremens, seizures, potentially FATAL |
| Benzodiazepines (prolonged use) | Seizures, potentially FATAL |
| Opioids | Very uncomfortable |
| Tobacco/cannabis/stimulants | Uncomfortable |
The scientific frameworks, the planned models, the protocols of a future AI assistant, the success metrics, and the project's making-of. To understand how it's built.
Planned grid, not yet in the app. Today, nothing is measured automatically: no watch or band reading, no voice analysis. The app relies on what the person reports. The Source and Thresholds columns below are design targets, not current features. Intended principle: assess the craving's intensity, type, and context to tailor the response.
| Dimension | Source | Indicators | Thresholds |
|---|---|---|---|
| Physiological intensity | Wearable (HRV, heart rate) | Low HRV + elevated heart rate | HRV < baseline -20% |
| Self-reported intensity | Check-in / mood picker | Visual scale (not numeric) | Warm color = urgency |
| Timing | History + clock | Time, day, known patterns | Pattern match >70% |
| Sleep context | Wearable (night) | Duration, regularity, awakenings | Disrupted 2+ nights = ↑ vulnerability |
| Emotional context | Voice journal + NLP | Prosody, keywords | Detects anger/sadness/anxiety |
| Social context | Self-reported | Isolation, conflict, event | Isolation >48h + weekend |
| Craving type | Pattern analysis | Automatic vs. emotional vs. social | → response technique |
Positive weather. Optional micro neuro-module. Preventive activity suggestion.
Weather signals risk ahead. Proactive protection plan. "Your sleep was disrupted and it's Friday night. Want us to get ready?"
Guided SOS. Live craving timer. Personal crisis kit. Emergency contact if enabled.
"I used" button → zero judgment. It happens, nothing to prove, nothing to hide. Context is logged with the entry, with no extra screen or question about the moment, and nothing is collected without the person's knowledge: everything stays reviewable and editable in the Journal. "OK. We're here. Debrief whenever you're ready."
Not yet in the app. Intended principle: not to pick ONE approach, but to combine them based on the craving's context.
| # | Approach | Researcher | Limbra application |
|---|---|---|---|
| 1 | Reward circuits | Nora Volkow (NIDA) | Predictive salience → Craving weather |
| 2 | Allostasis & stress | George Koob (NIAAA) | Allostatic load → My skills |
| 3 | Pleasure-pain balance | Anna Lembke (Stanford) | The "Dopamine isn’t pleasure" module (wanting ≠ liking) |
| 4 | Craving extinction | Judson Brewer (Brown) | Habit loop → SOS, interruption |
| 5 | Compulsive habits | Barry Everitt (Cambridge) | Habit → compulsion shift → Contextual Journal |
| 6 | Neuroplasticity & learning | Marc Lewis (Radboud) | Deep learning → Progress & replay |
| 7 | Social context & choice | Carl Hart (Columbia) | Social determinants → Trigger mapping |
| 8 | Computational craving 🇫🇷 | Marc Auriacombe (Bordeaux) | EMA modeling → Craving weather, ML |
20 founding references · 17 peer-reviewed journal articles · 3 books · NEJM, Nature, Lancet, PNAS · including French research · See the full page (in French) →
| Technique | Source | Duration | When |
|---|---|---|---|
| RAIN | Brewer/MBRP | 5 min | Active craving, level 2-3 |
| Body scan | MBSR/MBRP | 10 min | Prevention, stress |
| Cognitive defusion | ACT | 5-10 min | Intrusive thoughts |
| Values clarification | ACT | Weekly | Motivation |
| Pattern interrupt | Behavioral | A few seconds | Acute craving, level 3 |
| Breathing sync | Physiological | 2-3 min | Anytime |
| Craving replay | Metacognitive | 60 sec | Post-craving |
| Craving surfing | Brewer | 5-8 min | Moderate craving |
| Technique | Reason for exclusion |
|---|---|
| Streak counter | Shame cycle when it resets |
| Aversion / punishment | Reinforces shame, counterproductive |
| Chromotherapy | Pseudoscience, no causal evidence |
| Points-based gamification | Activates reward circuits = the problem |
| Generic positive visualization | Not evidence-based for SUD |
| Unsupervised hypnosis | No possible supervision in an app |
Planned tone for a future assistant, not yet in the app. The examples below are style targets, not messages the app shows today.
Not a therapist. Not a coach. Not a buddy. A scientific companion: someone who understands the person's brain, doesn't judge, and knows when to speak and when to stay quiet.
| Context | Tone |
|---|---|
| Prevention (calm morning) | Light, informative, brief |
| Alert (risk detected) | Direct, concrete, action-oriented |
| Active craving | Minimal, sensory, no unnecessary words |
| Craving weathered | Warm, acknowledging |
| After use | Neutral, curious, zero judgment |
| Suicidal crisis | Immediate redirection, no attempt at therapy |
| Metric | What it measures | Why it's better |
|---|---|---|
| Moments weathered | Active resilience | The person DID something |
| Sensitive windows identified | Self-awareness | The person understands their brain |
| Metric | Target |
|---|---|
| Retention D7 / D30 / D90 | >50% D7, >30% D30 (vs. ~15% for current apps) |
| SOS use during a craving | Measures real-world usefulness |
| CSAPA referral rate | The app moves people toward care |
| NPS | >50 = excellent |
"How a graphic designer builds an app in 2026" · A 100% transparent journal of the collaboration with Claude (Anthropic).
~55h Thomas • ~43h Claude • ~98h combined over several months
| Task | Ratio | Insight |
|---|---|---|
| Scientific research | 70/30 | Vision and curation are 100% human |
| Color palette | 65/35 | The designer decides the colors. Full stop. |
| Logo | 60/40 | AI = ultra-fast design intern |
| Design benchmark | 55/45 | The AI doesn't know what's "beautiful" |
| One-pager for addiction specialists | 30/70 | The AI writes well when the brief is clear |
| SOS prototype (web demo) | 25/75 | A non-developer can produce a working prototype |
| Landing page | 20/80 | The human 20% makes the difference |
| Design tokens | 15/85 | Pure technical work = AI's sweet spot |
"I built an app with AI in one weekend" → Bullshit. Limbra is ~100 hours of work over several months: an app available today on Google Play; content not final. AI doesn't eliminate the work, it transforms it: less time spent executing, more time spent deciding.
A project wires itself
like a brain.
Every reply creates a connection.
Lived experience, a clinical eye, plain curiosity, or a wish to track your own use: Limbra adapts to what it's taught. Testers, clinicians, organizations, press; come leave a trace.
Get involved (in French)Here are the questions I'm most interested in. No need to answer them all.