Thalassa Assistant

An assistant that answers while the conversation is still running.

Thalassa listens to your conversation and writes an answer into a window only you can see — grounded in your own documents, with the source shown.

What it does

It hears the question and has the answer written before you need it.

Thalassa follows a live conversation, notices when a question is aimed at you, and drafts an answer while the other person is still speaking. The window it writes into is excluded from screen sharing and screen recording — the people you are talking to do not see it. Before the conversation it builds a briefing from your own documents. Afterwards it writes a report: what was decided, which objections came up, what to follow up on.

Where your data is

Your knowledge base stays on your machine. We could not read it if we wanted to.

Documents are read, indexed and searched locally; only the passages that match a question travel with the request. One exception at ingest, which we would rather name than have you discover: a scanned PDF contains pictures of text, so each page is sent to our backend in Frankfurt once, to be turned into text. It happens when you add the file, never during a conversation, and it can be switched off. Conversations are not stored. Live translation is optional and sends the other side's speech to DeepL while the meeting runs — leave it off and nothing leaves the European path (privacy policy, section 3.6).

What it does not do

It does not record, and it does not decide what you are allowed to use.

It does not record your meetings. It does not decide whether you are allowed to use assistance in a given setting, and it cannot know the rules that apply to you — that judgement stays yours. Where a selection process forbids aids, using it is not permitted; that stands in the terms, not just here.

Across the whole conversation

Where the time actually comes back

The gain is not only in the minutes on the call. It is in the preparation that otherwise costs you an evening, and in the follow-up that otherwise never happens.

7

Hours back every week

Modelled on 15 meetings a week — preparation, live answers and follow-up taken together. Your own number depends on how you work.

Before

Preparation that is already done

A full briefing built from your own documents
Rehearsal — you practise against the hard questions before they arrive
Meeting analysis: what it is about, who is across the table, where it will get tight

During

Backup nobody can see

Answers from your knowledge base — sourced, not guessed
Where needed, from verified official sources, with the citation
Live translation when you and your counterpart share no language
Out of your screen share — with a self-test that proves it, not a promise

After

Follow-up that does not slip

A debrief with open points and the commitments you made
Analysis of how the conversation went — what landed and what did not
The transcript itself is deleted; what you need stays

Product

Built for the conversations you can't afford to get wrong

Grounded answers

First-person answers drawn from your own knowledge base first, with the internet only as a fallback. Predictive answering can start composing while the other side is still speaking.

Live translation

Follow a call in a language you don’t speak: the other side is translated for you in real time. Billed per session minute from prepaid credit (€0.23/min) — you only pay when you use it.

Speaks your languages

Transcription in 17 languages, answers in 16, and the interface in 8: German, English, Spanish, French, Italian, Japanese, Korean and Chinese. All three are set independently — read the app in German while the meeting runs in English and the answers come back in German.

Discreet overlay

Your answers stay on your own screen. On Windows the window is excluded from capture by the operating system; on macOS it is absent from single-window shares, and since macOS 15.4 Apple no longer lets any app opt out of a full-screen share — so the app says so, and a built-in self-test checks the real state before you start. Focus mode, adjustable opacity, auto-dim and a restore shortcut included.

Conversation modes & playbooks

Eight modes — from sales and negotiation to coding interview and legal — shape the structure of every answer. Eight ready-made playbooks anticipate the questions and objections of a conversation type.

Briefing, scenario & practice

Walk in prepared: a pre-call briefing built from your documents, a scenario lab for “what if they say X”, and a practice mode that interviews you — no audio needed.

Screen analysis

A chart in the meeting, a task in the editor, a table in the deck: one click captures your screen for a single answer. The capture is used once and never stored.

Confidential mode

Names, file references and addresses are replaced with stable placeholders before anything reaches the language model, and translated back before the answer reaches your screen. If the anonymiser is unavailable, there is no answer — never a silent pass-through — and the minute is credited back.

After the call

An optional summary, a final report with open points and commitments, and pinned session notes — while the transcript itself is deleted. Your history keeps metadata only.

Also on board

Knowledge-base profiles per client
OCR for scanned PDFs
Only what changed is re-indexed
Auto-answer & adjustable sensitivity
Answer style & length controls
Custom instructions about you
Google sign-in
Single-meeting tickets — no subscription
Student plans
Auto top-up with a spending alert
Windows & macOS

Where the line runs

What Thalassa does during the call — and where every answer comes from.

More than a note-taker

Note-takers earn their keep from the archive. Thalassa answers while you are still speaking.

A note-taker records the call and hands you a summary afterwards. Thalassa works in the middle of it: it listens, searches your own documents and puts the answer, with its citation, into a window only you can see. Three things stay fixed. It does not join the meeting as a participant — there is no bot in the attendee list. Your knowledge base is searched on your own machine and never uploaded, and after the call no recording and no transcript is left behind. And the assistant window is your private window beside the conversation, not a concealed tool: it stays out of your screen share, and how far that holds on which operating system is set out plainly on the privacy page.

The difference, in plain terms

Two questions: where does the answer come from — and what is left afterwards?

A window only you can see, and listening through system audio instead of joining as a bot: the neighbouring tools have caught up on both. The difference sits in two places. First, where the answer comes from — your own documents and cleared sources, with a citation, rather than a model's general memory. Second, what is left afterwards: nothing. No recording, no archive, no analysis of your conversations. The first is a feature; the second is a business model — and nobody can copy it without giving up their own.

Answers with a source, not from memory

Your 2026 price list is in no model. It is in your knowledge base.

A general language model answers from what it has read. It has never seen your framework agreement, nor your case file — and it answers anyway, plausibly and wrongly. Thalassa draws its answer from two named bodies: your own documents, and a fixed set of verified legal and regulatory sources — official texts from the jurisdictions the capability matrix sets out. Every answer drawn from them names the act, the article, the level and the version date. And because Thalassa hears the question itself, nobody has to type while someone is talking to them.

Your documents first

Citations you can check.

Answers are drawn from your uploaded documents — product specs, price lists, case files. The internet is only a fallback, and when sources conflict, your knowledge base wins. Every answer names the document and the page it came from — no unverifiable claims. MCP connects compatible tool providers in every plan. You choose which tools are available and confirm every proposed call with the exact inputs it will use. Results can feed into answers and briefings; the provider is shown. The existing import of text resources remains available. Supported MCP connections.

Questions people actually ask

The things worth knowing before you try it.

Who builds Thalassa?

Thalassa Consulting Kft., Budapest. Full company details are in the imprint.

Can you read my documents?

No. They are read, indexed and searched on your machine. What reaches us is the passage that matches a question, for the length of that request — plus the one named exception for scanned PDFs, which you can switch off.

Do you store my conversations?

No. There is no archive and no transcript on our side. That is not a setting; it is how the product is built — which is also why it cannot offer you a searchable history later.

Where does the backend run?

Frankfurt, Germany (AWS eu-central-1). The one path that leaves it is optional live translation via DeepL, and only while it is switched on.

Is it allowed where I want to use it?

That judgement is yours. Thalassa cannot know the rules of your employer, your client or your examination board, and it does not pretend to. Where aids are forbidden, using it is not permitted.