Private AI agent: what makes an agent private, and how Evanesca's works
A private AI agent is an AI that carries out a task for you — searching the web, reading your files, writing documents — without keeping, reading or reusing what you hand it. Evanesca is a private AI agent and chat: in its Agent mode, every call to the model goes to a model run in Europe, the task and its files are not kept by Evanesca, no code is ever run, and spending stops at the cap you set.
Facts checked on 6 Nov 2026
What an AI agent is
A chat answers a question. An agent takes a task — “compare these three quotes and draft a reply”, “find my rights on this broken washing machine and make me a one-page PDF” — and works through it in steps. At each step the model reads everything so far, decides what to do next and calls a tool: search the web, read a file, write a document, create an image. The tool's result goes back to the model, and the loop starts again until the task is done how tool calling works.
That loop is why privacy is harder for an agent than for a chat. One task sends your documents to the model many times over, passes queries to a search engine, writes files somewhere, and may run for minutes while you are not watching each step. Every one of those places is a place where your data could be kept.
What makes an AI agent private
Four questions separate a private AI agent from one that only says so.
- Where does each part run? The loop that drives the task, the model it calls at each step, the search engine and the image generator can be four companies in four countries.
- What does it keep? The task, its intermediate steps, the files it reads and writes, and the logs around all of them. A private agent keeps none of the content, and says what it keeps around it.
- What can it touch? An agent with your email, your accounts, a browser that signs in for you, a payment card or a code interpreter can do more — and can be steered by a web page into doing more than you asked. The fewer doors it has, the smaller the risk.
- Who can read it, and does it train? As with any private AI: the provider's staff, the operator of the model, and whether your tasks can become training data.
The full checklist, for chats as well as agents, is on the private AI page.
Evanesca's Agent mode: a private AI agent, step by step
- You give it a task: the result you want, the documents or images it needs, a model that can call tools, and a spending cap — €0.50, €1, €2, €5 or €10, €2 by default. Agent mode works with a model you choose: Auto mode does not run tasks.
- It makes a plan: for a task of several steps, it writes a to-do list and ticks items off as it goes.
- It looks things up: it rereads the files of the conversation and, only if you have switched web search on, searches the web and reads pages.
- It produces: text files, Word and PDF documents, Excel workbooks that hold values only, and images made with FLUX.1 [schnell].
- It hands the work over: an answer, with its steps, searches and files underneath, each file ready to download.
It acts only through 11 tools, and every one of the 16 models Evanesca serves was measured calling tools, with a real call. How Agent mode works, tool by tool.
Where each part of a task runs, and what is kept
| Part of the task | Where it runs | What is kept |
|---|---|---|
| The agent's loop: the steps, the file and document tools | Evanesca's server, at Vercel — a US company — in its Paris region privacy | No copy of the task or its files: held in memory while it runs. A record without content: model, cap, status, number of web calls, start and end times |
| Each call to the model | In France by Scaleway and OVHcloud (as a fallback), in Mistral AI's own EU region, in Germany by Verda and AKI.IO and in Switzerland by Infomaniak | The rule of the model's operator, below |
| Web searches and pages read | Linkup, a French company — only if you switch web search on | The query the model writes and the addresses of the pages; never the conversation. Linkup's own policy applies source |
| Images | Infomaniak, in Switzerland (outside the EU, under an EU adequacy decision) | Infomaniak states that queries are “neither recorded nor used to train models” source |
| Your files and the conversation | Your browser | Until you delete the conversation; sent to the server at the start of each task, for that task only |
On the model side, each operator states its own rule: OVHcloud keeps “only the data required for billing purposes” source; Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days source; Mistral AI keeps no input or output beyond the time needed to answer, as Evanesca's account has been on zero data retention since 17 November 2026 source, and lets its customers object to training — Evanesca's account objects source source; AKI.IO says prompts and outputs are neither logged nor stored, and keeps usage metadata for 90 days source source; Verda says it stores no request and holds the context in an in-memory inference cache removed within 2 hours source; Infomaniak states that queries are “neither recorded nor used to train models” source.
So the honest description is a private AI agent whose model calls all stay in Europe — not an agent run end to end by European companies: the loop runs on a US company's servers in Paris, which keep no copy of the task. The passport under the answer names the model, the operator and the country of each call; sovereignty explains who controls each operator.
What the agent cannot touch
- Code. No interpreter, no sandbox, no terminal: a script it writes stays a text file, a workbook holds values only, and an HTML or SVG file it creates is downloaded, never opened in the page.
- Your accounts and your computer. It has no access to your email, your accounts or the files on your computer — only to the files of the conversation and, with web search on, to public web pages. It fills in no form, signs in nowhere and buys nothing.
- Your memory profile. Memory does not apply in Agent mode: the agent works from its own instructions and tools.
- More money than you allow. Before each call to the model, Evanesca reserves the most that call could cost; if it does not fit in what is left of the cap or of your balance, the call does not go, the task stops, and the files written so far are kept. The cap holds to the precision of that worst case, so the total can end slightly above it.
- More steps or time than set. 40 steps at most, the last one without tools, and no new step after 10 minutes. With web search on, 40 web calls per task and 300 per 24 hours per account, at €0.0076 a search and €0.0015 a page read; and 12 images per task at most.
Fewer doors also means fewer ways to be misled: a web page the agent reads cannot make it send an email, move money or run a program, because it has no tool for any of them.
Compared with ChatGPT and other AI assistants
ChatGPT offers more agent features than Evanesca — deep research, scheduled tasks, Codex, and an always-on agent on its Pro plan source. What differs is what happens to your words. On OpenAI's plans for individuals, as its pages read on 6 October 2026:
| Question | ChatGPT | Evanesca Agent mode |
|---|---|---|
| Where conversations are stored | “On OpenAI systems and our trusted service providers’ systems in the US and around the world” source | In your browser; the models run in Europe sovereignty |
| How long they are kept | Until you delete them, then up to 30 days source | Not kept by Evanesca: in memory while the task runs privacy |
| Training on them | Possible unless you opt out; feedback can bring the whole conversation in source | Never by Evanesca; Mistral AI lets its customers object, and Evanesca's account objects |
| Who may read them | Authorised OpenAI staff and service providers, for abuse, support, legal matters, or model improvement unless you opted out source | No one at Evanesca; the model's operator, under its own rule |
| How you pay | Free plan with limits; Go, Plus and Pro subscriptions source | Per use, VAT included, within the cap of each task; credit from €5, no subscription required pricing |
Where ChatGPT is the better choice: more agent features, OpenAI's own frontier models, voice, and apps on the web, iOS and Android source. For anything that involves no private or confidential information, it offers more. The full comparison is on Evanesca vs ChatGPT; whether a European equivalent exists is answered on European ChatGPT.
Frequently asked questions
What is a private AI agent?
A private AI agent is an AI that carries out tasks for you — searching the web, reading your files, writing documents — without keeping your data, letting others read it or using it to train a model. To judge one, ask where each part of the task runs, what it keeps, and what it can touch.
Is Evanesca's Agent mode private?
Its model calls go only to models run in Europe, the task and its files are not kept by Evanesca, and Evanesca never trains on them. The loop that drives the task runs on Evanesca's server at Vercel, a US company, in its Paris region, with no copy kept. The operators that run the models have their own rules: Scaleway, for instance, may keep requests that cause errors or look abusive for up to 14 days.
Can the agent run code or use my accounts?
No. It has no interpreter, sandbox or terminal, and no access to your email, your accounts or the files on your computer. It works only with the files of the conversation and, if you switch web search on, public web pages. It fills in no form and buys nothing.
How much does an AI agent task cost on Evanesca?
You set a cap for each task — €0.50, €1, €2, €5 or €10, €2 by default, never more than your balance. Each call to the model is billed at the model's price; with web search on, a search costs €0.0076 and a page read €0.0015. A subscription is optional.
Is a private AI agent the same as a local AI agent?
No. A local agent runs on your own computer, with a model small enough for its memory. Evanesca's Agent mode runs in the cloud, calls models run in Europe, and keeps no copy of the task.
Read next
- The best ChatGPT alternatives, by what you needThe best ChatGPT alternative depends on your need: privacy, European law, free use, research, code or offline. Claude, Lumo, Duck.ai, Vibe and more.
- Private AI: what it means, and how to judge onePrivate AI keeps your words from being stored, read or used for training. The three kinds, a checklist to judge any of them, and the main options compared.
- Is there a European ChatGPT? What it means, and the optionsIs there a European ChatGPT? What to check — EU company, servers in Europe, the CLOUD Act — and how Mistral Vibe, Lumo and Evanesca compare.
- European AI: who makes it, who runs it, and how to use itWhat European AI covers: models made in Europe, European operators and data centres, the AI Act — which ones Evanesca runs, and how to use them today.
- Sovereign AI: what the word covers, and how to check a claimSovereign AI means control of the model, the servers, the operator and the law. A checklist to judge any claim, and where Evanesca stands on each layer.
- AI for lawyers and law firms: what it is good for, and how to keep client secretsAI for lawyers: drafting, summaries, translation, contract review — and client confidentiality. What the French bar says, a checklist for law firms.