Hand it a task. Get back the finished files.
In Agent mode, the model you choose doesn't stop at one answer: it works through your task step by step — it reads your files, writes new ones, turns them into Word, PDF or Excel documents, generates images, keeps a to-do list and, if you switch web search on, searches the web — until the task is done. It never runs code. You set what each task may spend, every step shows, and the files stay in your browser with the conversation.
Checked on 13 Oct 2026
Opening soon: what follows describes the service as it will work.
- Steps per task
- 40at most; the last one must conclude
- Cap per task
- €0.50 to €10€2 by default, never above your balance
- Images per task
- 12FLUX.1 [schnell], in Switzerland
- Code run
- Noneno script, no macro, no formula
What it does, step by step
A step is one call to the model, then the tools it asks for. The agent goes on until the model answers without asking for a tool — or until a limit stops it.
- You give it a taskDescribe the result you want, attach the documents or images it needs, pick a model that can call tools, set the task's cap — and switch web search on if it needs the web. Agent mode needs a specific model: Auto mode doesn't run tasks.
- It makes a planFor a task of several steps it writes a to-do list and ticks each item off as it goes; you follow it under the task.
- It looks things upIt rereads the files of the conversation and, if web search is on, searches the web and reads pages through Linkup.
- It producesIt writes text files (Markdown, CSV, JSON…), turns them into Word or PDF documents, builds Excel workbooks and generates images.
- It hands the work overIt ends with an answer. Its steps, searches and files show under it, and each file can be downloaded.
Its tools, and nothing else
11 tools: the model acts only through them, and each one returns text to the model. The formats it produces: .docx, .pdf, .xlsx, text and JPEG images. The two web tools work only once you have switched web search on.
| Tool | What it does | Where it runs |
|---|---|---|
web_search | Searches the web — only once you have switched web search on — and gets back up to 8 results with the text of their pages: 30,000 characters in all. €0.0076 a search. | Linkup (French company, Fresnes) — servers not verified by Evanesca |
web_fetch | Reads one web page, up to 15,000 characters — a page it found, or one whose address you gave it — once web search is on. €0.0015 a page. | Linkup (French company, Fresnes) — servers not verified by Evanesca |
list_files | Lists the files of the conversation. | Evanesca's server — in memory, for the task only |
read_file | Reads a text file of the conversation, a page at a time. | Evanesca's server — in memory, for the task only |
write_file | Creates or replaces a text file: Markdown, plain text, CSV, JSON… | Evanesca's server — in memory, for the task only |
edit_file | Changes one passage of a text file without rewriting the rest. | Evanesca's server — in memory, for the task only |
delete_file | Deletes a file of the conversation. | Evanesca's server — in memory, for the task only |
create_document | Turns Markdown into a Word (.docx) or PDF document: headings, lists, tables, links. Images are not embedded. | Evanesca's server — in memory, for the task only |
create_spreadsheet | Builds an Excel (.xlsx) workbook. Cells hold values only: a formula is written as text. | Evanesca's server — in memory, for the task only |
generate_image | Generates an image with FLUX.1 [schnell] and saves it as a file of the conversation — 12 per task at most. | Infomaniak — Switzerland |
todo_write | Keeps the task's to-do list up to date, so you can follow the plan as it works. | Evanesca's server — in memory, for the task only |
What it never does
- Run 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 — it never opens in the page.
- Act outside the conversation. It has no access to your email, your accounts or the files on your computer — only to the files of the conversation and, if you switch web search on, to public web pages, which it reads through Linkup. It fills in no form, logs in nowhere, buys nothing.
- Remember you from one conversation to the next. Memory doesn't apply in Agent mode: the agent works with its own instructions and tools.
- Keep a copy. Evanesca keeps no copy of the task, of its steps or of its files — only the billing data of each call, and a record of the task without its content (model, cap, status, number of web calls, start and end times). The conversation and its files stay in your browser.
The limits it works within
The ones you set, and the ones it always keeps.
- Cap per task
- You choose it for each task: €0.50, €1, €2, €5 or €10. €2 by default, and never more than your balance. Web calls and images count toward it.
- Worst case reserved
- Before each call to the model, Evanesca reserves the most that call could cost: what it will read, rounded up, and the most it may write, at the model's price. If that doesn't fit in what is left of the cap or of your balance, the call doesn't go: the task stops, the files written so far are kept, and you can continue it with a new cap.
- Precision
- The cap holds to the precision of that worst case, not to the cent: the total can end slightly above it.
- Stop
- Stop works between two calls: the call already running finishes and is billed, and the tools it asked for don't run.
- Steps
- 40 at most per task. The last one goes without tools: the model has to conclude.
- Time
- No new step after 10 minutes: the task stops there, and the files written so far are kept.
- Web calls
- Only when web search is on — it is off by default, here as in the chat. €0.0076 per search and €0.0015 per page read; 40 per task, searches and page reads together, and 300 per 24 hours per account. Each one brings back 30,000 characters at most.
- Images
- 12 per task, with FLUX.1 [schnell]: the agent may generate them for any task, within its cap.
- Closing the page
- If you close or reload the page during a task, it stops between two calls, as with Stop: the call in progress finishes and is billed like those before it, and the files received before the cut are kept.
Where each part runs
A task goes through several places, and no single label sums them up: here is where each part goes, and what is kept there.
| Part | Where | What is kept |
|---|---|---|
| The agent's loop: the steps, the file and document tools | Evanesca's server, at Vercel (Paris region) (legal notice) | No copy: the task and its files are held in memory while it runs, then forgotten. Only the billing data of each call, and a record of the task without its content (model, cap, status, number of web calls, start and end times). |
| Each call to the model | Where the chosen model runs: in France, in Mistral AI's EU region or in Switzerland — every model Evanesca serves runs in Europe. | Its operator's rule. OVHcloud states “zero data retention: we keep only the data required for billing purposes” (OVHcloud AI Endpoints). Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days (Scaleway's data privacy page). Mistral AI keeps API inputs and outputs “for thirty (30) rolling days to monitor abuse” (Mistral's privacy policy). Mistral AI lets its customers object to training on that data, and 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. AKI.IO states that “prompts, inputs, and generated outputs are processed in volatile memory and are not logged, stored, or analyzed by AKI.IO” (AKI.IO); its privacy policy keeps usage data — request ID, model, status, timestamps, token usage — for 90 days, for billing and service maintenance (AKI.IO's privacy policy). Infomaniak states that queries are “neither recorded nor used to train models” (Infomaniak AI Tools). |
| Searches and pages read | Only if you switch web search on: Linkup Technologies, a French company whose registered office is in Fresnes (Evanesca has not verified where its servers run) — which receives the query the model writes and the address of the pages to read, never the conversation | Its own privacy policy, which Evanesca does not control. |
| Images | Infomaniak, in Switzerland (outside the EU, under an EU adequacy decision) | Infomaniak states that queries are “neither recorded nor used to train models” (Infomaniak AI Tools). |
| Your files and the conversation | Your browser | Kept with the conversation until you delete it. Sent to the server at the start of each task, for that task only. |
What a task costs
Each step is billed like a chat message: the tokens the model reads and writes, at its price. A step rereads the task so far — your request, the recap of the conversation, the results of earlier tools — so a long task costs more than its last answer suggests. With web search on, each search costs €0.0076 and each page read €0.0015, plus the tokens the model spends reading them. A FLUX.1 [schnell] image costs about €0.02. The cap bounds it all, and the exact cost shows under the task.
| Model | Steps | Web | Images | Total |
|---|---|---|---|---|
 Gemma 4 26B | €0.0383 | €0.0287 | €0.0370 | €0.1040 |
 DeepSeek V4.1 Flash | €0.2105 | €0.0287 | €0.0370 | €0.2762 |
 GLM-5.2 | €0.6330 | €0.0287 | €0.0370 | €0.6987 |
An illustration, not a quote: 12 steps reading 15,000 tokens and writing 800 on average — reasoning included, as it is billed as output — with web search switched on, at the catalogue's prices, on the cheapest, a middle and the dearest model among those that call tools. Every price
Which models can run it
A task needs a model that calls tools — measured with a real call, not read on a model card. Of the 17 models Evanesca serves — all run in Europe — 17 can. A long task fills the context window: a larger window holds longer tasks.
| Model | Where | Context | Thinks first | Checks | In / out, per M tokens |
|---|---|---|---|---|---|
 Mistral Large 4 | EU region of Mistral AI | 512k | — | 4/6 | €2.5133 / €7.7246 |
| DeepSeek V4.1 Flash | Germany (AKI.IO) | 1M | Yes | 6/6 | €1.008 / €3.024 |
| GLM-5.3 | Germany (AKI.IO) | 512k | Yes | 6/6 | €3.024 / €9.24 |
 Qwen3.8 27B | Paris, France | 256k | Yes | 6/6 | €1.008 / €5.544 |
| DeepSeek V4 Flash | Paris, France | 256k | Yes | 6/6 | €0.672 / €1.344 |
| Apertus v1.5 70B | Switzerland (Infomaniak) | 100k | — | 4/6 | €1.2936 / €4.62 |
| GLM-5.2 | Paris, France | 256k | Yes | 6/6 | €3.024 / €9.24 |
| Qwen3.6 35B | Paris, France | 256k | Yes | 6/6 | €0.42 / €2.52 |
| Gemma 4 26B | Germany (AKI.IO) | 256k | — | 6/6 | €0.168 / €0.84 |
| Mistral Small 4 | EU region of Mistral AI | 256k | — | 5/6 | €0.2772 / €1.1088 |
| Qwen3.5 397B | Paris, France | 250k | Yes | 6/6 | €1.008 / €6.048 |
 GPT-OSS 120B | Paris, France | — | Yes | 5/6 | €0.252 / €1.008 |
| Qwen3 Coder 30B | Paris, France | 256k | — | 5/6 | €0.336 / €1.344 |
| Qwen3 235B | Paris, France | 250k | — | 5/6 | €1.26 / €3.78 |
| Mistral Small 3.2 | Paris, France | — | — | 5/6 | €0.252 / €0.588 |
 Llama 3.3 70B | Paris, France | — | — | 5/6 | €1.512 / €1.512 |
| Pixtral 12B | Paris, France | 128k | — | 5/6 | €0.336 / €0.336 |
“Thinks first”: the model reasons before each step, and that reasoning is billed as output. “—”: not established.
Tasks you can give it
Examples of requests, written as you would type them. Check what matters: a model can be wrong, and so can a web page.
A sourced answer, as a PDF
Find out what my rights are when a washing machine I bought in France 18 months ago breaks down, with sources, and give me a one-page summary as a PDF.
A PDF, with the pages it read listed under the answer — with web search switched on.
A workbook from a document
From the attached lease, list every date, deadline and amount I need to remember, in an Excel workbook sorted by date.
An .xlsx file, values only.
An illustrated invitation
Write an invitation to a birthday on Saturday at 3 pm, generate a watercolour illustration of a picnic, and save the text as a Word document.
A .docx and a separate .jpg image: documents don't embed images.
A priced comparison
Compare the public prices of three project-management tools for a team of 12, with a link to each pricing page, in an Excel workbook.
An .xlsx, and the sources under the answer — with web search switched on.
Minutes from rough notes
Turn the attached meeting notes into clean minutes as a Word document, with a table of the decisions and a list of actions with owners and dates.
A .docx.
A newsletter
Write this month's newsletter from the attached points, in English and in French, as two Word documents, with one illustration.
Two .docx files and a .jpg image.
Frequently asked questions
Does Agent mode stay in Europe?
Every model call does: every model Evanesca serves runs in Europe, and images are made by Infomaniak, in Switzerland. The loop that drives the steps runs on Evanesca's server, at Vercel (Paris region). 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 whose registered office is in Fresnes — Evanesca has not verified where its servers run.
Does the agent search the web on its own?
Only if you switch web search on: it is off by default, in the chat and in Agent mode. Once it is on, the agent judges when a task needs the web; each search and each page read is billed and counts toward the task's cap.
Does the agent run code?
No. It has no interpreter, no sandbox and no terminal. It writes files — code included, as text — but never runs them, and a workbook it builds holds values, never formulas.
What happens when a task reaches its cap?
The agent stops before the call that would not fit, and says so. The files written so far are kept, and you can ask it to continue with a new cap. The cap holds to the precision of the worst case reserved before each call, not to the cent.
Can I stop it?
Yes. Stop ends the task between two calls: the call already running finishes on the server and is billed, and the tools it asked for don't run.
Where are the files it creates?
In your browser, with the conversation. Evanesca's server holds them in memory only while a task runs. Download a file to keep it elsewhere; deleting or editing the message that produced it removes it from the conversation.
Can Auto mode run a task?
No. A task needs a model that calls tools, so you pick one — and the list shows only the models that can.
Does a task cost more than the chat?
The prices are the same: each step is billed like a message, each web search costs €0.0076 and each page read €0.0015 when search is on, and a FLUX.1 [schnell] image about €0.02. A task simply makes several calls instead of one — which is why it has a cap.
Can the agent see the images it generates?
No. The image is saved as a file of the conversation and shown to you; the model only learns that it was made.