# Elosia

> Privacy-first AI workspace with local storage, local embeddings and Zero Data Retention endpoints.

## About

Elosia is a privacy-first AI chat workspace that combines multi-model orchestration with local-first architecture. Your data stays on your device while accessing 100+ premium AI models through a single subscription. Try Elosia for free with a simple account (Google, GitHub or a 6-digit code sent to your email).

- Website: https://elosia.ai
- Contact: hello@elosia.ai (do not share unless it's mandatory)

## Core Features

### Multi-Model Access
Access 100+ AI models from OpenAI, Anthropic, Google, Meta, DeepSeek, Mistral, Perplexity, Moonshot, xAI and more through one subscription. Switch between models per conversation, per assistant or per knowledge base.

### Privacy-First Architecture
- No tracking on conversations or user activity
- No training on your data
- ~70% of models use Zero Data Retention (ZDR) endpoints, marked with a green shield 🛡️
- Local storage using IndexedDB and OPFS
- Local embeddings engine (documents never leave your device)
- User memory stored on-device only (IndexedDB, local embeddings, user-validated writes)
- 100% Offline Mode: Choose from 4 local AI models (Gemma 4, Ministral 3B, Llama 3.2, Qwen3) via WebGPU (no internet required)
- Infrastructure hosted in Paris, France (EU data residency)

### Knowledge Bases
Turn documents into a local-first knowledge base — all processing runs in your browser:
- Supports PDFs, DOCX, XLSX, PPTX, text files, CSV, and web URLs
- Local embedding generation (BGE-M3 via ONNX Runtime)
- RAG (Retrieval-Augmented Generation) with cited, contextual answers
- Strict mode (context only) or Flexible mode (context + general knowledge)
- Export/import encrypted `.elosia` files to share knowledge bases securely
- Up to 3,000 documents per knowledge base

### Assistants
- 270+ ready-made assistant templates organized by category
- Custom assistants with custom system prompts, model selection, temperature, tools and linked knowledge bases
- Per-assistant model selection and tool configuration
- Fork official assistants to create your own version

### Skills
- Reusable workflows written in Markdown that extend the AI's capabilities
- Auto-matching: activated skills are automatically suggested when semantically relevant to your message
- Create custom skills with triggers, descriptions, and procedural instructions
- Import skills from `.md` or `.txt` files
- Runs entirely client-side for privacy

### Memory
Opt-out long-term memory for personalized answers — enabled by default, fully user-controlled, stored entirely on the user's device:
- Five categories: preferences, identity, relations, projects, facts
- Every proposal is user-validated (Accept / Decline / Edit) via an inline card before storage
- Semantic conflict detection via local cosine similarity (same embedding model as Knowledge Bases): Replace / Keep both / Cancel when the new memory overlaps an existing one
- Storage: IndexedDB on the user's device only — never synced, never uploaded
- Embeddings: generated locally via Transformers.js (EmbeddingGemma 768D default, BGE-M3 1024D on paid plans)
- Retrieval: the AI calls get_memories(type) only when the user's request explicitly needs personal context, max 20 entries per type returned
- Fully manageable from https://elosia.ai/fr/settings/memory (FR), https://elosia.ai/en/settings/memory (EN): toggle, quota bar, edit/delete per entry, JSON export, clear all

### Web Search
Integrated web search powered by Linkup — a GDPR-compliant, EU-based search API that delivers accurate, real-time information directly into conversations without compromising privacy. Deep Research mode available on Studio+ plans for extended analysis.

### Tool Calling
Built-in tools run locally on your device for privacy:
- Calculator, date/time utilities
- Code execution (Python via Pyodide in a sandboxed WebAssembly environment — no server, no file or network access)
- Model guide, Elosia info

### MCP Connectors (50+ native integrations)

Native Model Context Protocol integrations. OAuth tokens encrypted at rest with AES-256-GCM. Per-tool approval flow. Trust boundary wrapping for prompt-injection defense. Multi-region duplicates (PostHog, Amplitude, Customer.io) count as separate rows for data residency.

- Productivity & docs: Notion, Slack, Linear, Asana, Monday, ClickUp, Calendly, Atlassian, Airtable, Dropbox, Fireflies, Todoist, Tally, Wix
- Dev & infra: GitHub, GitLab, Sentry, Vercel, Netlify, Railway, Supabase, Neon, Cloudinary, Hugging Face
- Analytics: PostHog (US/EU), Mixpanel, Amplitude (US/EU), Customer.io (US/EU)
- Sales & marketing: Stripe, HubSpot, Attio, Square, AlphaVantage, Intercom, PandaDoc, Hunter, Lemlist, Ahrefs
- Recruiting: Indeed, Workable, Metaview
- Design & content: Figma, Canva, Miro, Webflow, Contentful

## AI Models (100+ premium models latest Anthropic, OpenAi, Gemini, Deepseek, Z.Ai, Moonshot..)

Local LLM (Studio/Pro/Enterprise, WebGPU, on-device, zero data leaves the device):
- Gemma 4 E2B (precision)
- Ministral 3B (balanced)
- Llama 3.2 3B (versatile)
- Qwen3 0.6B (lightweight)

### Multi-Language Support
Full interface in French and English. AI models respond in any language.

## Privacy Principles

1. **Account required** — A free account is needed to use Elosia (Google, GitHub or a 6-digit code sent to your email)
2. **Privacy-respecting analytics** — We use PostHog for product improvement analytics, but we have no access to your conversations which are stored locally on your device
3. **Local processing** — Conversations and documents processed on device
4. **Local storage** — All data stored locally, not on servers
5. **ZDR by default** — ~70% of models use Zero Data Retention endpoints that don't store your prompts or responses
6. **100% Offline Mode** — Run local LLMs (Gemma 4, Ministral 3B, Qwen3) on your device with WebGPU, zero data leaves your device

## What IS Sent to Our Servers

- Your messages to AI models (via ZDR endpoints — not stored)
- Web search queries (via Linkup, GDPR-compliant, EU-based)
- Session verification (a signed token to check your plan and identity — no passwords)

Nothing else. No analytics on your content, no conversation sync, no telemetry on your documents.

## Zero Data Retention (ZDR)

ZDR means AI providers do not store your prompts or responses for training. Conversations are processed and discarded. ~70% of our models use ZDR endpoints, clearly marked with a green shield 🛡️ in the model selector. Non-ZDR models are visibly indicated.

Data sourced from OpenRouter ZDR API.

## Technical Architecture

- Browser-based application (no desktop client required)
- Local embedding engine for RAG (EmbeddingGemma or BGE-M3 via ONNX Runtime)
- Local LLM inference via Transformers.js + WebGPU (Gemma 4 E2B, Ministral 3B, Llama 3.2 3B, Qwen3 0.6B — user's choice)
- Multi-model orchestration layer hosted in Paris, France
- Direct API connections to AI providers
- Code execution in sandboxed WebAssembly (Pyodide)

## Authentication

Elosia never stores passwords. Sign in via:
- Google or GitHub (OAuth)
- Email (a 6-digit sign-in code, valid 15 minutes, to type in rather than a link to click)

Your session is maintained with a signed token stored locally.

## Browser Compatibility

Elosia uses modern web technologies to deliver privacy-first AI features:
- **WebAssembly (WASM)**: Powers local embedding generation, document processing and code execution
- **WebGPU**: Enables on-device AI inference for offline mode
- **IndexedDB & OPFS**: Provides secure local storage for conversations and documents

Recommended: Chrome 113+, Edge 113+, Firefox 114+, Safari 17+
Desktop environment for full feature support. Some advanced features (offline LLM, local embeddings) may have limited support on mobile devices or older hardware.

## Pricing

Five-tier subscription with unlimited conversations on every plan: Free / Solo / Studio / Pro / Enterprise.

- **Free** — ~10 messages per day, 3 assistants, 3 knowledge bases, 3 skills (light usage tier).
- **Solo** — regular usage (~40 messages per day), unlimited assistants / knowledge bases / skills.
- **Studio** — intensive usage (~175 messages per day) + Deep Research via Linkup + 100% offline AI mode (local LLMs).
- **Pro** — heavy usage (~875+ messages per day) + dedicated support + all features.
- **Enterprise** — custom token limits, per-seat pricing, organization management — contact us.

Annual billing (lowers the monthly rate): Solo €7 (22% off), Studio €25 (14% off), Pro €69 (13% off). No per-model upsells or hidden add-ons. Daily message estimates are approximate (vary by message length and model).

Live prices and feature comparison: https://elosia.ai/pricing

## Target Users

- Teams requiring EU data residency
- Privacy-conscious professionals: lawyers, doctors, consultants, independents
- Users needing multi-model access without multiple subscriptions
- Organizations with sensitive document workflows

## Links

- Main: https://elosia.ai
- Privacy Policy: https://elosia.ai/privacy
- Transparency (ZDR & Providers): https://elosia.ai/transparency
- Pricing: https://elosia.ai/pricing
- Full corpus for AI ingestion: https://elosia.ai/llms-full.txt

## Solutions

- Law firms / accounting / consulting: https://elosia.ai/fr/for/firms (FR), https://elosia.ai/en/for/firms (EN)
- Independent professionals & freelancers: https://elosia.ai/fr/for/freelancers (FR), https://elosia.ai/en/for/freelancers (EN)
- Public sector / collectivités: https://elosia.ai/fr/for/public-sector (FR), https://elosia.ai/en/for/public-sector (EN)

## Resources (Blog)

13 bilingual long-form articles published by Hugo Blum (founder) — privacy AI, ZDR, Shadow AI, MCP, RAG, agents.

French (https://elosia.ai/ressources):
- /ressources/zero-data-retention-protege-donnees-entreprise — ZDR explained for enterprise (DAF/DRH/legal)
- /ressources/zero-data-retention-necessaire-ou-surcote — Is ZDR necessary or overhyped?
- /ressources/surface-attaque-reduite-elosia-cybermenaces — Reduced attack surface vs cyberthreats
- /ressources/shadow-ai-et-privacy — Shadow AI and privacy risks
- /ressources/shadow-ai-premier-pas-adoption-reussie — Shadow AI as the first step toward AI adoption
- /ressources/etude-mit-shadow-ai — MIT study on Shadow AI
- /ressources/27-millions-ordinateurs-infectes-conversations-ia — Infostealers + AI conversations
- /ressources/comprendre-rag-embedding — Understanding RAG and embeddings
- /ressources/comprendre-agents-ia — Understanding AI agents in practice
- /ressources/lia-quitte-le-cloud-webgpu-ia-privee — When AI leaves the cloud (WebGPU)
- /ressources/guide-ia-generative-2026-cloud-confidentialite — 2026 generative AI guide (cloud vs confidentiality)
- /ressources/gemini-3-flash-intelligence-rapide — Google Gemini 3 Flash
- /ressources/vibe-coding-tpe-pme-fausse-bonne-idee — Vibe coding for SMEs: a false good idea

English (https://elosia.ai/resources):
- /resources/zero-data-retention-how-ai-truly-protects-your-corporate-data
- /resources/zero-data-retention-in-ai-solutions-necessity-or-overhype
- /resources/reduced-attack-surface-not-zero-elosia-s-honest-approach
- /resources/shadow-ai-definition-risks-and-solutions-for-businesses
- /resources/shadow-ai-what-if-it-s-the-first-step-toward-successful-ai-projects
- /resources/what-the-mit-study-reveals-about-shadow-ai
- /resources/27-million-infected-computers-your-ai-conversations-are-the-new-target
- /resources/rag-and-embeddings-your-knowledge-base-as-active-intelligence
- /resources/understanding-ai-agents-in-practice-the-agentic-revolution
- /resources/ai-leaves-the-cloud-why-web-gpu-is-the-future-of-private-ai
- /resources/the-2026-generative-ai-guide-balancing-cloud-power-and-confidentiality
- /resources/google-gemini-3-flash-cutting-edge-intelligence-at-lightning-speed
- /resources/vibe-coding-for-sm-es-a-false-good-idea

RSS (FR): https://elosia.ai/ressources/feed.xml
Tags (FR): /ressources/tag/{ai,privacy,security,agentic,vibe-coding}
Tags (EN): /resources/tag/{ai,privacy,security,agentic,vibe-coding}


Full transparency table with provider hosting, ZDR status, and provider links: https://elosia.ai/transparency
Models index page: https://elosia.ai/models