Terms of service and sale.

In short

• Evanesca is a private AI agent and chat whose models all run in Europe, with an API for developers; it is published by FLEECE AI SASU, a French company.
• No subscription: you use prepaid credit; you add credit from €2; credit never expires, not even if you close your account.
• Prices are in euros, VAT included, per model and per million tokens; the exact cost of each answer shows under it.
• In Agent mode, you set a spending cap for each task — €2 unless you choose another — and you can stop the task between two steps.
• If you are a consumer, you can withdraw from each credit purchase within 14 days.
• Evanesca keeps no copy of your messages. Answers, documents and images made by an AI can be wrong: check what matters.

The service and these terms

These terms govern the use of Evanesca — the agent and the chat in your browser, the OpenAI-compatible API and Business workspaces — between you and FLEECE AI SASU, 10 rue Pierre Poisson, 63400 Chamalières, France (details in the legal notice). They apply from the opening of the service, and you accept them when you create your account.

If you use Evanesca for an organisation, you accept them on its behalf. A Business workspace may add a signed agreement and a data processing agreement (DPA); where they differ from these terms, they prevail.

Your account

You sign in with your email address: we send you a link. An account is personal; keep access to your mailbox safe.

You are responsible for what is done with your account and with your API keys. Keep keys secret, and revoke one at once if you think it has leaked.

You can close your account at any time, from your account space or by writing to us from its address — see Suspension and closure.

Credit and prices

Evanesca runs on prepaid credit, in euros. There is no subscription, no recurring payment and no card kept for later.

You add credit when you choose, from €2. Credit never expires, not even when you close your account: what you have not used stays attached to the account, and you find it again if you come back (see Suspension and closure).

Each message is charged for the tokens the model processes — what you send and what it writes, reasoning included, counted as output — at the price per million tokens shown for that model on the models page when you send it. Prices include our margin and VAT. The exact cost shows under every answer.

The chat's other features are charged from the same credit. Each step of an agent task is charged like a message. An image is charged per second of generation — €0.00924 a second today, so about €0.02 for an image from FLUX.1 [schnell] and €0.05 for a portrait from PhotoMaker V2. Dictation is charged per second of audio — about €0.005 a minute. Web search, if you switch it on, is billed per call: €0.00756 per search and €0.001512 per page read, on top of the input tokens the model spends reading the results; through the MCP server, only the price of the call is billed, since your client's own model reads the results. The chat also makes small charged calls of its own: a title for each new conversation, follow-up questions after a complete answer, a summary when a long conversation is compressed and, with memory on, the profile update — their cost shows in the account's usage history, and the follow-ups' cost also under the answer. With memory on, the profile is also charged as input with each message. Prices are on the pricing page.

When your credit runs out, new messages are refused until you add credit. An answer already being written is finished and charged, which can take your balance slightly below zero; the difference is taken from your next credit purchase.

Prices can change — for instance when an operator changes its own. The price that applies is the one shown when you send the message.

Payment and invoices

Credit purchases are paid online, in euros, VAT included, through Stripe (Stripe Payments Europe, Limited, Ireland). The amount you pay is the amount credited: nothing is added at checkout.

Stripe issues an invoice with every credit purchase — a PDF, with VAT shown — and emails it to you automatically; the email confirming the purchase also links to it. Business workspaces receive VAT invoices.

Right of withdrawal (consumers)

If you are a consumer, you have 14 days from each credit purchase to withdraw from it, without giving a reason (French Consumer Code, art. L221-18).

Before you pay, you are asked to request that the service start at once, so that you can use the credit straight away. If you then withdraw, you pay for the credit already used and we refund the rest (art. L221-25); if that request was not collected, we refund the whole credit purchase. Refunds are made within 14 days of your withdrawal, by the same means of payment (art. L221-24).

How to withdraw, and the model withdrawal form: right of withdrawal.

Acceptable use

You agree not to use Evanesca:

• to break the law or infringe the rights of others — for example to create or spread illegal content, to harass someone or to deceive;
• to impersonate someone, or to create sexual content involving a minor, in any form;
• for a practice the AI Act bans (Regulation (EU) 2024/1689, art. 5);
• to attack, probe or overload the service, the operators' systems or other websites — through the agent and the web search too —, or to get around its limits;
• to resell access to the service or the API without our written agreement.

Images follow the same rules, and these in particular:

• add a face photo only if you have the right to use it: your own, or that of a person who agreed;
• never make an image of a real person without their consent, and never a sexual or intimate image of a real person, even with their consent;
• never make a deceptive deep fake: an image made to pass for a real person, place or event, in order to mislead.

Generated images are AI-generated content. They carry a machine-readable mark saying so — IPTC metadata in the file (AI Act art. 50(2)) — which a screenshot or a re-encode can remove. See the AI Act page.

Evanesca is not intended for decisions about people that the AI Act treats as high-risk: recruiting or selecting candidates; decisions about workers — promotion, dismissal, allocation of tasks, monitoring or evaluating their performance; assessing creditworthiness; or pricing life and health insurance (Annex III, points 4 and 5). We specify that Evanesca is not to be turned into a high-risk AI system (art. 25(2)).

We may suspend an account that breaks these rules — see Suspension and closure.

Answers, documents and images generated by AI

Answers, the documents the agent writes and generated images are produced automatically by AI models made by third parties — the maker of each model is named on the models page. Answers are labelled as AI-generated. All of them can be wrong, incomplete, out of date or biased, even when they look sure: check anything that matters before relying on it.

An answer about health, law or money is not a professional's advice. For a medical question, see a doctor.

We claim no rights over what you send or over what the models produce. Some models' licences set conditions on how their output may be used; each model's licence is on the models page.

If you use Evanesca at work, the AI Act asks you to say that content was generated by AI when you publish it: text meant to inform the public on matters of public interest, unless a person reviewed it and holds editorial responsibility for it; and an image that resembles real people, places or events and could pass for authentic — a “deep fake” (art. 50(4)). See the AI Act page.

Agent mode

In Agent mode, the model you choose carries out a task for you step by step, with tools: it can search the web and read pages, if you have switched web search on; list, read, write, edit and delete the files of the conversation; write Word and PDF documents from Markdown, and Excel spreadsheets without formulas; generate images, 12 per task at most; and keep a to-do list. It never runs code. Web search follows your switch, off by default: while it is off, the agent searches nothing on the web and reads no page. It may make images when the task calls for them, within its cap. A task takes 40 steps at most, and starts no new step after 10 minutes.

For each task, you choose a spending cap of €0.50, €1, €2, €5 or €10; without a choice, it is €2. The cap can never exceed your balance. Before each call to the model and each image, the most that call could cost is set aside from what is left of the cap; when what is left cannot pay for the next call, the task stops. Each step is charged like a chat message, each web call — a search or a page read — at its price, and each image like any other image — see Credit and prices. Web calls and images count toward the task's cap.

“Stop” stops the task between two calls: the call already running finishes, and it is charged.

The task runs on Evanesca's servers, and every model it calls runs in Europe (see Models and where they run). If web search is on, the agent's searches and page reads go to Linkup; its images go to Infomaniak.

The files stay in your browser, with the conversation: if you delete the conversation or your browser's data, they are gone, and we cannot recover them. Evanesca keeps no copy of what the agent reads or writes.

You are responsible for what you ask the agent to do and for what you do with what it produces: check a document before you rely on it or send it. We may stop a task that breaks these terms or puts the service at risk.

Web search

Web search is off by default, in the chat and in Agent mode. If you switch it on, the model can search the web and read pages when your question needs fresh information, through Linkup, a French company: the queries it writes, and the addresses of the pages it reads, go to Linkup, never your conversation. Evanesca has not verified where its servers run; its own privacy policy applies. In a chat, each search brings back 8 results at most, each cut at 15,000 characters, and the model makes 4 rounds of searches at most before it answers. In Agent mode, the agent makes 40 web calls per task at most — searches and page reads together — and 300 per 24 hours per account; each call brings back 30,000 characters at most.

Search results are written by third parties. Evanesca does not check them: a page can be wrong, out of date, or written to mislead the model. Check the sources before you rely on what they say.

Each search costs €0.00756 and each page read €0.001512, VAT included, on top of the input tokens the model spends reading the results. Through the MCP server, only the price of the call is billed.

Images

Images are generated by FLUX.1 [schnell] or, for a portrait from 1 to 4 face photos, by PhotoMaker V2; they run at Infomaniak, in Switzerland. In the chat, image generation is off by default; once it is on, up to 4 images per message; through the API, up to 5 per request; in 1024×1024, 1792×1024 or 1024×1792.

The rules for face photos, real people and deep fakes are in Acceptable use.

Generated images can be inaccurate: faces, hands, text and details can come out wrong, and a portrait may not look like the person. They are AI-generated content — see Answers, documents and images generated by AI.

The licences of the image models — FLUX.1 [schnell]: Apache-2.0; PhotoMaker V2: Apache-2.0 (SDXL base: CreativeML OpenRAIL++-M) — may set conditions on how the images are used.

Dictation, read-aloud and memory

Dictation turns up to 60 seconds of speech into text, through Mistral AI, a French company (Voxtral Mini): the text appears in the message box, and nothing is sent until you send it. It is charged per second of audio. Mistral AI keeps the recording and its transcript 30 rolling days to monitor abuse; it lets its customers object to training on them, and Evanesca's account objects.

Read-aloud uses your device's own voices: it is free, and the text never leaves your device.

Memory is off until you switch it on. Then a profile of 4,000 characters at most, kept in your browser, is updated after each answer by a small charged call, and sent with each message. You can read, change or erase it at any time.

Models and where they run

Every model Evanesca serves runs in Europe — in France (Scaleway, OVHcloud as a fallback), in Mistral AI's EU region, in Germany (AKI.IO) or in Switzerland (Infomaniak); Switzerland is outside the EU, under an EU adequacy decision. If the operator of a model fails, your message fails; it is never sent elsewhere, and no model call leaves Europe. In Auto mode, Evanesca picks the model among those of the catalogue, from signals in your message, at no extra cost. The sovereignty page describes every route.

Each operator applies its own rule to what it receives; the privacy policy quotes them. Mistral AI, which runs Mistral Large 4, Mistral Small 4 and dictation, keeps API inputs and outputs 30 rolling days to monitor abuse, and lets its customers object to training on them; 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.

Operators update and retire models. We may add, replace or withdraw a model; the models page always shows the current catalogue, and each answer's passport names the model the operator actually used.

Your data

Evanesca keeps no copy of your messages. Your history — with the agent's files, the images and your memory profile — stays in your browser: if you clear it or lose your device, it is gone, and we cannot restore it. How we handle personal data is set out in the privacy policy; Business workspaces also get a data processing agreement.

Availability

We do our best to keep Evanesca available, but we do not promise an uninterrupted service: maintenance, an outage at an operator or at a provider such as Linkup, or a model's retirement can interrupt it. No availability level is promised outside a signed Business agreement. We may change or improve features.

Liability

If you are a consumer, we are liable to you as the law provides, and nothing in these terms limits that.

If you use Evanesca for your business, our total liability for a claim is limited to the credit you bought in the 3 months before it, and neither party is liable for indirect loss — lost profits, lost data, business interruption — except where the law forbids such limits.

In every case, you are responsible for what you send and for what you do with the answers, the documents and the images.

Suspension and closure

We may suspend or close an account that breaks these terms. Unless there is an ongoing attack or a legal obligation, we tell you first and give you the chance to explain.

You can close your account at any time, from your account space or by writing to us from its address. Closing it does not make your credit expire: you find it again if you come back with the same address. For that, if you have credit left, we keep a fingerprint of your address, the former account's identifier and its balance. Everything else is deleted, except what the law requires us to keep — the details and the periods are in the privacy policy. This credit is not refunded, except under the right of withdrawal or where the law requires it.

Changes to these terms

We may update these terms when the service changes. Changes that matter are announced by email and on this page at least 14 days before they apply. If you do not accept them, you can close your account before that date.

Applicable law and disputes

These terms are governed by French law. If you are a consumer living in another country of the European Union, you keep the protection of the mandatory rules of your country.

If something goes wrong, write to us first. Before the service opens, we will name here the consumer mediator you can turn to free of charge if we cannot agree (French Consumer Code, art. L612-1).

Failing an amicable solution, the competent courts decide; as a consumer, you can also go to the court of the place where you lived when the contract was made.

Contact

Questions about these terms: contact@fleeceai.agency. By post: FLEECE AI — Evanesca, 10 rue Pierre Poisson, 63400 Chamalières, France. The publisher's details are in the legal notice.