NutriCoach.AI

A multi-agent AI nutrition and fitness coach for real food and real results. 7 LangGraph ReAct agents turn your profile into calorie targets, a natural-protein Indian (or American) meal plan, a walking or minimal-equipment gym plan and realistic deadlines. Daily tracking, streaks and a local body illustrator keep you going.

• Works with any LLM. OpenAI, Claude, Gemini, DeepSeek or a free local model via Ollama. There is also a no-AI mode. Every number is computed in code and the model only writes the explanations, so even a 1-4 GB local model gives a correct plan.
• Real food only. Protein is 1.3 g/kg from eggs, chicken, fish, paneer, tofu, dal, curd, milk, soya and nuts. No whey, protein bars or Greek yogurt. Budget mode uses only the cheapest sources.
• Built for Indians, at home or abroad, with Asian BMI cut-offs and Indian meals. American food is one click away.
• Walking for everyone. Gym users get a push/pull/legs plan using only dumbbells, a bench and a pull-up bar, plus bodyweight moves.

Screenshots

Agents + 3 effort paths	Today: checklist, meals, workouts
Body illustrator (light theme)	Plan + live BMI gauge

Features

3 ways to your goal	Easy, Medium and High effort paths, each with calories, steps, gym days, kg/week, kg/month and a real deadline date. Unsafe targets are flagged (loss is capped at 1% of body weight a week).
Daily plan	A tickable checklist for food, protein, water, steps, gym and sleep. 7 days of meals with portions, and snacks to eat before meals so you don't overeat.
Weight gain	Lean-gain surplus, homemade high-protein shakes and calorie-dense Indian foods, with treats allowed twice a week.
Activity	Step targets with a 6-week ramp-up and easy runs on the High path. Gym users get PPL or full-body sessions with animated exercise diagrams.
Supplements	Bone health and recovery only: vitamin D3, calcium, magnesium, B12/omega-3 for vegetarians, optional creatine.
Tracker	Log calories eaten and steps. Each day is labelled a fat loss, fat gain or maintenance day.
Gamified	Walking, healthy-eating, calorie-budget and gym streaks (2+ sessions a week), XP levels from Rookie to Legend, motivation quotes.
Body illustrator	Cartoon "now" and "goal" bodies drawn on your device from body-fat % and muscle % sliders. No AI, no upload.
Metrics	BMI gauge (Asian or WHO cut-offs), obesity class, estimated body fat, waist-to-height ratio, healthy weight range, BMR and TDEE.
Profiles	Up to 3 saved profiles, each with its own plan and history.
Themes	Hacker green (default), light, dark and neon pink/purple. Responsive from 320 px phones to 4K, in portrait and landscape.

Agent graph

BMI Agent ─► Nutrition Agent ─► Critic ──(score < 7, max 2)──► Nutrition Agent
                                  │
                                  ▼
                 Meal + Supplement Agent ─► Activity Agent ─► Plan Critic ──(fails)──► Meal Agent
                                                                   │
                                                                   ▼
                                                              Coach Agent ─► END

Agent	Tools	Model tier
BMI	calculate_bmi, classify_risk	small
Nutrition	compute_tdee (3 paths), build_macro_targets	small
Critic	score_plan + code checks (numbers stated, no banned foods)	large
Meal + Supplement	build_meal_plan, get_supplement_stack	small
Activity	walking_plan, gym_plan	small
Plan Critic	rule checks: calories, protein, food repeats, safe rate, movement	large
Coach	safety_check, synthesise_report	large

Why it works on weak models: planner.py computes the whole plan first, and each agent's prompt carries the exact numbers. If a model can't call tools, the bare model still writes the text. If it fails or writes nothing, the planner's own text fills in. The critics then check the text against the numbers and the banned-food list.

Compared with popular apps

	NutriCoach.AI	HealthifyMe	MyFitnessPal	Noom	Fitbod
Free, no account	✅	Freemium	Freemium	❌	❌
Runs offline with a local AI	✅	❌	❌	❌	❌
Indian meal plans	✅	✅	Partial	❌	❌
Natural protein only (no whey)	✅	❌	❌	❌	❌
Realistic deadline per effort level	✅	Partial	❌	Partial	❌
Dumbbell + bench gym plan	✅	Partial	❌	❌	✅
Streaks and levels	✅	✅	Partial	✅	Partial
Body-composition visual	✅	❌	❌	❌	❌
Data stays on your device	✅	❌	❌	❌	❌

Setup

python -m venv venv
venv\Scripts\pip install -r requirements.txt
venv\Scripts\python.exe app.py        # http://localhost:5000

Open Settings to pick a provider and paste a key, or choose Local LLM (Ollama, free):

1. Install Ollama from ollama.com.

2. Pull a model that fits your RAM. Settings lists them smallest first:

   Model	Download	RAM
   qwen2.5:1.5b	1.0 GB	4 GB
   llama3.2:3b	2.0 GB	8 GB
   qwen2.5:7b	4.7 GB	16 GB

   ollama pull llama3.2:3b

Use a small model for the narrating agents and a larger one only for the critics and the coach. Any newer model id can be typed in.

Keys, settings, profiles and your log are stored only in ~/.nutricoach/, never in the repo.

Desktop app (.exe)

venv\Scripts\python.exe desktop.py    # chromeless app window
build_exe.cmd                          # one-file NutriCoachAI.exe via PyInstaller

Project structure

agents.py       7 ReAct agents, their tools, the LangGraph graph
planner.py      rule engine: metrics, paths, meals, activity, checks, fallback text
llm.py          provider layer (OpenAI, Anthropic, Gemini, DeepSeek, Ollama, offline)
tracker.py      profiles, daily log, streaks, levels, quotes (SQLite)
app.py          Flask API + SSE stream of the graph
desktop.py      desktop window launcher
templates/      single-page UI
static/css, js  themes, app logic, art.js (body illustrator + exercise diagrams)
docs/           screenshots and preview image

Disclaimer

NutriCoach.AI is for education and general wellness. It is not medical advice. Talk to a doctor or registered dietitian before changing your diet or starting supplements, especially if you have a medical condition.

License

MIT