Skip to content

Evanesca opens soon: this page shows how the chat will work.

Why Evanesca

Eight reasons to switch, each with its proof.

Evanesca is a private AI agent and chat — a European alternative to ChatGPT. It runs AI models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google in Europe, searches the web, writes documents and creates images, and keeps no copy of your messages. Here is what that means in practice — and where to check each claim.

Checked on 29 Sep 2026 · Mistral AI, Infomaniak and Linkup pages read on 13 Oct 2026

Opening soon: what follows describes the service as it will work.

The reasons, and their proof

01

An AI agent that does the task

Give it a task, not just a question: in Agent mode the model writes Word, PDF and Excel documents, generates images and, if you switch web search on, searches the web and lists its sources, step by step, until the task is done — within a spending cap you set (€2 by default), and with no code ever run. Its files stay in your browser.

Proof: Agent mode · Web search · Image generation

02

Your messages never leave Europe

Every model Evanesca serves runs in Europe — in France by Scaleway (Paris) and OVHcloud (Gravelines, as a fallback), in Mistral AI's EU region, in Germany by AKI.IO or in Switzerland by Infomaniak (outside the EU, under an EU adequacy decision). No model call leaves Europe: not as a backup, not when servers are busy. None of these companies has a US parent, so they are out of the CLOUD Act's direct reach. Servers in Europe are not enough on their own: a US company stays subject to the CLOUD Act wherever its servers are, so Evanesca also looks at who controls the operator. Web search is off by default; if you switch it on, the queries the model writes, and the addresses of the pages it reads, go to Linkup, a French company — Evanesca has not verified where its servers run; its own privacy policy applies.

Proof: Sovereignty · What the CLOUD Act says · Web search

03

No copy of what you send

Your prompt and the answer stay in memory only while the answer is written; your history and the agent's files live in your browser. About your messages, Evanesca keeps only billing data — tokens and cost — and never uses them to train AI. The operators have their own rules, always stated: Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days; Mistral AI keeps its API's inputs and outputs for 30 days to monitor abuse — that covers Mistral Large 4 and Mistral Small 4 and dictation — and lets its customers object to training: Evanesca's account objects, except for what is sent to Mistral Large 4: it is a preview, and Mistral's commercial terms let it use the data of preview models for training, whatever the objection; Linkup, which runs the web searches when you switch them on, follows its own privacy policy.

Proof: Transparency · The answer passport

04

A passport under every answer

Each answer says which model wrote it, who ran it and where, under which law, and what was kept — no content. Every model shows its checks before you use it, and a missing check is printed on the passport, not hidden.

Proof: The passport, field by field

05

Open models, their checks in plain sight

Models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google, all run in Europe. Every model shows its checks before you use it — Licence, Provenance, Security, GDPR, Benchmarks and Live test — and the ones it misses are printed, not hidden. Of the 17 models Evanesca serves, 16 have open weights: the same model can run elsewhere, with no lock-in. More models will join when a European operator runs them.

Proof: Every model · Benchmarks

06

You pay for what you use

No subscription: credit from €2, each purchase with its invoice. Credit never expires, not even if you close your account. Prices include VAT, and the exact cost of each answer shows under it. A web search costs €0.0076, a FLUX.1 [schnell] image about €0.02 in generation time, and every Agent task has its cap.

Proof: Pricing

07

Ready for teams

A Business workspace adds a signed Art. 28 DPA, the sub-processor list with notice before any change, SSO and roles, spending limits on top of each task's cap, VAT invoices and passport exports.

Proof: Business

08

One account, chat and API

The same models through an OpenAI-compatible API — chat completions, the model list and embeddings — at https://evanesca.app/api/v1, with your key in an Authorization: Bearer header. Code written for OpenAI's SDK only needs a new base URL and a new key.

Proof: Developer docs

What you can trust it with

Contracts

Summarise a contract, compare two versions, list the clauses to renegotiate: the model that reads it runs in Europe, and Evanesca keeps no copy.

HR and payroll

Draft a letter, summarise a policy, check a payslip. Not for deciding about people: the AI Act treats that use as high-risk.

Sourced research

Hand the agent a question that needs the web, with web search switched on: it searches, reads and gives you a memo as a PDF, its sources listed. Check the ones that matter.

Code

Debug, review or explain code: Evanesca keeps no copy of it.

Figures before they are public

Analyse results, a budget or a deal memo before they are announced: Evanesca keeps no copy.

Illustrations

An image for a post or an invitation, made in Switzerland for about €0.02. Never a picture of a real person without their consent — and never an intimate one, even with it.

What Evanesca does not promise

  • That operators keep no data at all: Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days; Mistral AI keeps its API's inputs and outputs for 30 days to monitor abuse; OVHcloud keeps the data billing requires; AKI.IO keeps usage metadata — request IDs, timestamps, token counts — for 90 days.
  • That Linkup's side is covered by Evanesca's promises: if you switch web search on, the queries the model writes go to Linkup, which follows its own privacy policy; Evanesca has not verified where its servers run.
  • That Agent mode is one place: its loop runs on Evanesca's server, each model call goes to its model's operator — always in Europe — and searches, if you switch web search on, go to Linkup.
  • To be compliant on your behalf: no company can certify itself GDPR-compliant. So Evanesca publishes what it does, rule by rule.
  • That an answer is right: a model can be wrong, and so can a web page it read. Check what matters — the passport tells you which model wrote what, and the sources which pages it read.

Frequently asked questions

What makes Evanesca different from a typical AI chat?

Four things: it carries out tasks, not just answers — with web search, documents and images; where your messages are processed (in Europe: every model Evanesca serves runs there); what is kept (no copy at Evanesca); and what you can check (a passport under every answer). The comparison with ChatGPT, Gemini, Copilot, Claude and Meta AI is on the home page, with its sources.

What can the agent do?

Write files, create Word, PDF and Excel documents and images, keep a to-do list and, if you switch web search on, search the web and read pages — step by step, within the cap you set. It never runs code. How Agent mode works.

Who publishes Evanesca?

FLEECE AI (SASU), a French company based in Chamalières, SIREN 932 329 188. Legal notice.

Can I check these claims myself?

Yes. Every fact about a third party is dated and linked to its source: the sources page quotes the policies behind each comparison, and every model's page links to its official model card.