Korai's web chat pays with blind tokens, so a token does not tell Korai who you are or when you got it. That does not make you invisible: the computer that answers reads your message while it answers, and our web host sees your IP address. This page explains what is protected and what is not.
Every time a European asks ChatGPT, Claude, or Gemini a question, their words cross the Atlantic, pass through data centers subject to the Cloud Act, and leave identifiable traces tied to their credit card. That's not inevitable. It's an architectural choice that can be undone.
Korai bets on the opposite: open AI models (Qwen, Gemma and the next ones), running on computers lent by their owners, reached through anonymous digital tickets. When you chat on the web, those tickets are spent, not your identity: Korai cannot match a spent ticket to the claim that produced it.
We believe sovereign AI isn't decreed in Brussels. It's built, line of code by line of code, computer by computer, ticket by ticket.
Korai owns no data center. The network is made of computers lent by their owners. Each machine hosts an open model and answers the questions sent its way. The network is young: today one computer serves all requests.
No host has authority over the others. A central orchestrator, run by Korai in Paris, sends each request to one of them.
Your question travels over TLS on every hop. The orchestrator and the computer that runs the model read it in clear while answering. Korai's servers do not store the text of your messages.
Every request passes through the orchestrator before it reaches a host. It reads the request to route it and keeps no copy of its text: it routes, it relays the answer, it publishes what the network can do.
The orchestrator owns no graphics card and stores no model weights. On every request, it picks the right host among the connected computers based on the model requested and the capabilities the host advertises (memory, quantization, context length).
The reply streams from the host to your browser through the orchestrator. The stream is not encrypted end to end, so the orchestrator can read it while it passes. It does not store it.
The orchestrator continuously publishes the list of models each host loads based on its available GPU memory, with hot-swap updates and no worker restart.
In anonymous mode, no code joins an anonymous token to an account. Automated tests fail the build if a table or a query links the two sides.
In Paris, on Fly.io. It stores account data, usage counters and anonymous-token records, never the text of your messages.
In anonymous mode, nothing links a token to an account. This section states what that protects and what it does not.
Billing and anonymous tokens live in the same database, in separate tables with no foreign key and no SQL join between them, and tests check it. The question “what were this subscriber's conversations” has no answer in our data, because we do not store conversations.
Separate tables, no possible join.
Your browser gets tickets that Korai signs without seeing them (Privacy Pass). When you chat, a ticket is spent, not your name. Korai cannot match a spent ticket to the claim that produced it.
Your tokens are encrypted on your device, with a password only you know. The key never leaves your machine. Korai cannot open it.
Billing and anonymous-token records are in one Postgres database, in separate tables. No foreign key and no SQL statement joins them.
Every hop uses TLS: your device to our web host, the web host to the orchestrator, the orchestrator to the host computer. TLS is not end-to-end encryption: each of those servers reads the message in clear.
Every model Korai serves has publicly downloadable weights. No proprietary model runs on the network.
No outside security firm has audited Korai yet. Until one does, treat the guarantees on this page as our own statements, not certified ones.
Said plainly, so you can decide what to type.
While it answers, the message is in clear on the orchestrator and on the computer that runs the model. Treat it as a message sent to a service you do not control.
Our web host sees it on every request. Anonymous mode does not hide it; a VPN does.
Anonymous chats from the same browser share a random id, so they can be linked to each other (not to you).
With web search on, which is the default, the queries the model writes go through Korai's search server to Bing News, Stack Overflow, Wikipedia and Wikidata.
No host runs in a hardware enclave (TEE), and no relay or mixnet is in use.
Tus documentos se cifran en tu navegador antes de guardarse, y la indexación se hace en tu dispositivo. Cuando un espacio se usa en una conversación, los pasajes más relevantes se envían en claro al orquestador de Korai y a la GPU que responde.
Korai only ever serves open-weight models: publicly downloadable weights, commercial license, documented training. No proprietary model (GPT, Claude, Gemini) ever runs on our network. The fleet is audited every 30 days against public benchmarks (ArtificialAnalysis, Open LLM Leaderboard, Chatbot Arena). When a better model comes out, it's added and rolled out network-wide without human intervention.
Qwen 3.8-27B
Alibaba · Apache 2.0
Qwen 3.8-27B Thinking
Alibaba · Apache 2.0
Gemma 4 26B A4B
Google · Apache 2.0
Kimi K3
Moonshot AI · modified MIT
Korai relies on Rucio, the data-management system born at CERN, for the logistics of its model fleet, as part of the CERN Venture Connect program.
Rucio is the scientific data-management system built at CERN for the Large Hadron Collider: roughly one billion files, over one exabyte of data, across a network of around 120 computing centers. It's also been adopted outside particle physics, for example by LIGO and SKA.
Our fleet is made of quantized open-weight models, 15 to 30 GB each, refreshed every month, distributed to hundreds of partner hosts. Korai relies on Rucio to orchestrate that logistics at network scale.
Rucio orchestrates model replacement through declarative rules: a model enters the catalog, Rucio places it on the right hosts and retires the old one.
Models and evaluation sets are pre-positioned by Rucio as close as possible to the hosts that will use them.
Rucio posts end-to-end checksums to the audit log, giving regulated-sector clients a verifiable proof of integrity.
The same box as ChatGPT or Claude, with anonymous mode by default. Every feature below is live on korai.one. Available right now, no account or credit card needed.
Token-by-token responses, multi-turn, keeps conversation context. Compatible with existing OpenAI SDKs via /v1/chat/completions.
Drag and drop a PDF, a Word file, a .csv, .json, .md, .txt. Korai extracts the text in your browser and injects it into the prompt. 10 MB limit per file.
For questions about current events, prices, recent releases: Korai triggers a search via SearXNG (a self-hosted European instance) and cites its sources.
Regenerate the last response, edit your own messages, delete a conversation turn. History stays in your browser.
Ask for a budget, a tracking sheet, a structured data file. Korai generates a downloadable multi-sheet .xlsx in one click. 100% browser-side conversion.
Letters, reports, memos, meeting notes as Word or PDF (A4 rendering). Formatting preserved: headings, bold, italics, bullet lists.
Ask for a pitch deck, a summary, a training deck. Korai generates a .pptx with separate slides, titles, bullet points.
Ask for a calculation, an analysis, a chart. Korai writes Python executed locally via Pyodide (in-browser, no server). matplotlib, numpy, pandas available.
Breaks a task into a plan, executes the steps one by one (Python + file generation + web search), and summarizes at the end. Capped at 8 turns to stay controllable.
Plug in Linear, GitHub, Notion, Slack, or any MCP HTTP server. Configuration lives in your browser, your token stays local. One-click templates at /platform/integrations.
Lend your GPU to the Korai network. Earnings are tracked per request; payouts are not live yet. Cross-platform binary (Linux/macOS), self-connects, manages its own queue.
Your LangChain, LlamaIndex, or any OpenAI SDK integrations work by simply swapping the base URL. No proprietary Korai SDK to learn.
Blind cryptographic tickets (Privacy Pass / VOPRF P-384) issued from your browser, signed by Korai without ever being seen.
Korai's servers do not store the text of your messages. Your history stays in your browser unless you turn on sync.
Off by default. When on, your conversations are encrypted in your browser with a key only your devices hold. The server stores ciphertext.
The worker locks one copy of each request in RAM so it is not swapped to disk. That is best effort: other copies exist, and it does not stop the host's operator from reading the request.
The pipeline is ready on the chat side (image upload, OpenAI multimodal format). Missing a worker hosting a vision model (Qwen-VL, Gemma vision, Llama 3.2 Vision).
Text → image (FLUX, SD3, comparable open-weight models). Dedicated worker coming, always on a European worker.
Whisper for transcription, open-weight models for speech synthesis. All on a European worker, no third-party service.
A relay chain is built but not offered in the app yet. A mixnet would come later. Neither is in use today.
The wall between payment and chat is written into the SQL schema and checked by automated tests. Blind tokens add a cryptographic separation on top.
No proprietary black box. Every model is publicly auditable, its weights downloadable.
No conversation content is ever kept. The orchestrator only ever sees minimal technical metadata.
The orchestrator runs in Paris on Fly.io, the models on computers lent by their owners. The website is served by Vercel, a US company.
Hosts who lend their GPU earn in proportion to the service rendered: 75% of the user rate is credited to them. Payouts are not live yet.
Compliance is a design constraint, not an add-on: no message content is stored and data is kept to what the service needs.
Whether you're a curious developer, a lab with a GPU sleeping at night, or a public administration looking for an AI Act-compliant alternative: there's a place for you.