Free, model-agnostic AI nutrition tracking. Snap or describe a meal, and the AI model you choose returns estimated calories, macros, and portions — powered by affordable providers, especially Chinese LLMs.
// free on the chrome web store · bring your own key · your data stays yours
The name is the workflow. No forms, no barcode hunting, no database of foods to search.
Shoot a photo, paste an image from your clipboard, or just describe it in words. Whatever's fastest.
The AI model you chose returns estimated calories, protein, carbs, and fat — with its assumptions and a confidence score. Edit anything before you keep it.
It lands in your local log and updates your energy ring. Nothing left your device but that one call you asked for.
SnapBite picks the cheapest suitable model for the task — a vision model for photos, a text model for descriptions — then turns the reply into a clean, editable breakdown.
A meal photo, a pasted image, or a plain-text description of what you ate.
The model you chose recognizes the food and estimates portions from what it can see or read.
Returned as an editable card you can correct and save.
No forced provider, no DeepSeek subscription. Speaks the OpenAI-compatible and Anthropic protocols, so nearly any model works — and model IDs and endpoints are editable when providers change names.
Affordable-first, with Chinese LLMs like Qwen-VL, DeepSeek, and Kimi (Moonshot) leading. Use a low-cost text model to describe, a vision model for photos — you pick what fits your budget.
Photos need a vision-capable model; descriptions work with text models. SnapBite checks compatibility before it calls, so a text-only model on a photo gets a helpful nudge — not a silent failure.
Every meal lives on your own device. No account, no backend, no telemetry — the only network request is the one you trigger, straight to your chosen provider.
Everything is a visual estimate, never a medical measurement. SnapBite keeps its assumptions and a confidence score visible, and lets you correct any number before saving.
Free and MIT-licensed. Vanilla JavaScript, no framework, no build step — it opens instantly and stays out of your way.
SnapBite speaks the OpenAI-compatible and Anthropic protocols, so nearly any model works. Start at the top for the cheapest options — the starred picks are the Chinese-first recommended defaults.
| Provider | Best for | Vision | Get a key |
|---|---|---|---|
| Qwen-VL ★ | Chinese-first vision; trial quota | Yes | alibabacloud.com ↗ |
| DeepSeek ★ | Ultra-cheap text for the Describe tab | Text only | platform.deepseek.com ↗ |
| Kimi (Moonshot) ★ | Chinese-first, long-context; kimi-k3 reads photos & text | Yes | platform.moonshot.ai ↗ |
| SiliconFlow | Affordable vision + text models | Yes | cloud.siliconflow.cn ↗ |
| OpenRouter | One key, many providers & models | Yes | openrouter.ai/keys ↗ |
| Google Gemini | Fast models, generous free tier | Yes | aistudio.google.com ↗ |
| OpenAI | Reliable GPT-4o vision | Yes | platform.openai.com ↗ |
| Claude | Careful analysis with image support | Yes | console.anthropic.com ↗ |
| Grok | xAI vision where available | Yes | console.x.ai ↗ |
| Ollama | Local, offline & free on your own hardware | Model-dependent | ollama.com ↗ |
| Custom | Any OpenAI-compatible endpoint you run | Model-dependent | — |
// ★ recommended defaults · prefer nothing leaves your device? run a local model with Ollama
Choose a single provider, add its key in Settings, select a compatible model, and analyze your meal. That's it — no juggling accounts.
Photo mode needs a vision-capable model; Describe mode works with text models like DeepSeek. Model IDs and base URLs are editable, so you can update them whenever providers rename models or move endpoints.
SnapBite has no server of its own. There's nothing in the middle to log your meals, sell your habits, or hold your history hostage. Here's the entire data map.
Quick answers about the free AI calorie counter, supported models, privacy, and accuracy.
Yes. SnapBite is free and open source (MIT licensed). You bring your own AI API key, so the only cost is your chosen provider's usage — often a fraction of a cent per photo, and several providers (like Qwen, Gemini, and SiliconFlow) have free tiers.
Snap or paste a food photo and a vision AI model you choose identifies the foods, estimates the portions, and returns calories plus protein, carbs, and fat — with its assumptions and a confidence score. You can edit any number before saving it to your daily log. Prefer typing? The Describe tab estimates calories from a text description instead.
DeepSeek, Qwen-VL, Kimi (Moonshot), SiliconFlow, OpenRouter, Google Gemini, OpenAI, Claude, and xAI Grok — or a fully local model via Ollama. Model IDs and endpoints are editable, so you're never stuck when providers rename models. Chinese-first models like Qwen-VL, DeepSeek, and Kimi are the affordable defaults.
No sign-up and no account. You paste your own API key once, and your meal log lives locally on your device.
Yes. SnapBite has no servers and no analytics. Your meals stay on your device, and the only data that leaves is the photo or text you choose to analyze, sent directly to the AI provider you configured. With a local model via Ollama, nothing leaves your device at all.
Any Chromium browser that supports Manifest V3 — Chrome, Edge, Brave, and Arc.
They're visual estimates, not medical measurements. SnapBite keeps its assumptions and a confidence score visible and lets you correct any value before saving. It isn't medical advice.
Free on the Chrome Web Store — model-agnostic, open source, and yours to fork. Add it to Chrome and bring your own key.