My intelligent agent · EU-hosted

Your chief of staff, in the chat you already use.

MyIA reads your brief, chases what has gone quiet, and finds the thing in the file you forgot you had. It never sends anything without asking you first.

  • EU servers
  • Not used for training
  • Encrypted at rest
  • GDPR export & delete
MMyIA@MYIA_Davide_Bot · onlinetelegram
Anything on my plate tomorrow?9:14 AM ✓✓
Morning briefWed 13 Aug

10:00Call · A. Conti

12:30Lunch · M. Esposito

16:00Review · Q3 deck

3 followups waitingView →

9:14 AM
Draft Marco a nudge.9:15 AM ✓✓
Draft · ready to sendawaiting you

Ciao Marco — following up on the proposal you sent last week. Any chance you’ve had a chance to look it over?

SendEditCancel
9:15 AM

Nothing leaves the chat without an explicit approve.

Six workflows · Six habits

One small set of routines, learned once.

Not a chat-with-an-AI. A set of reliable moves the agent makes on your behalf — every one of them gated by your approval before anything leaves.

Morning brief

Every day at your local hour: today's calls, lunches, deadlines, and what has been waiting for a reply.

Fires at

07:30 localyour hour

Meeting prep

Before each call, a one-pager on the person, the company, and the last conversation.

Output

A. Conti · 10:00ready

Email draft

Reads the thread, drafts the reply, parks it for you to approve. Nothing sends without you.

State

draft · 142 wordsawaiting you

Calendar

Books the slot, warns you what it overlaps, and moves or cancels only after you approve.

Checks

Thu 15:00–16:00no clash

Follow-up chase

If a reply does not come, the agent nudges. You see every nudge before it goes out.

Waiting on

M. Esposito · 6 daysnudge drafted

Document Brain

Drop a PDF. Ask a question. Get the answer with a page citation. Delete the file and the agent forgets it.

Answer

contratto.pdf · p.4cited

Document Brain

Ask the file. The file answers back, with a page number.

Forward a PDF in chat or drop it on the dashboard. It is extracted, chunked, encrypted and indexed. Ask a question weeks later and the answer arrives with the passage it came from.

Read the technical deep-dive →
  • Hybrid retrieval

    Keyword search and meaning search, always both. A VAT number and "the thing about the deadline" are different questions.

  • Inline citations

    The answer names the file and the page. When the file does not say, it says that instead of filling the gap.

  • Injection defence

    A document is data, never instructions. Text that tries to give orders is wrapped, flagged, and written to your log.

Private by design

The whole reason this product exists.

MyIA was built because we wanted an assistant that did not read our email for its own purposes, train on our messages, or quietly take actions we would have to undo. Everything below is a commitment, not a setting you can forget to turn on.

  • EU residency

    Data lives in Falkenstein, Germany. There is no US region and there never will be.

  • Encrypted at rest

    Every file, every message, every chunk, under its own key. Deleting a file removes what the agent learned from it.

  • No training on your data

    By design, not a setting you can forget. Your messages, documents and memory train nothing.

  • An audit log you can read

    Every action, every model call, every approval. The log is yours to export and yours to delete.

  • Revocable grants

    Google, Telegram, WhatsApp — one click to revoke, and revocation is checked on the next read, not cached.

audit.log · yours to readlive

09:14:02chat.replyuser asked "what's on my plate tomorrow?"

09:14:02brief.morningfired at user's local hour

09:14:04llm.callthinking model · 1,204 tokens

09:14:06llm.calleveryday model · 312 tokens

09:14:07followup.track3 nudges queued · awaiting approval

09:14:09confirm.pendingemail to marco@… · awaiting user

09:15:21chat.replyuser asked "draft Marco a nudge"

09:15:22email.draftdraft composed · 142 words

09:15:24confirm.pendingemail to marco@… · awaiting user

09:18:44user.approvedsend authorized

09:18:45email.senddelivered · message id 18f2…9c

09:21:08document.uploadq3-deck.pdf · encrypted at rest

09:21:10retrieval.indexchunked · page anchors saved

09:23:55llm.callthinking model · 892 tokens

An illustration, not anyone’s real log — your own is in your account, and nobody else can read it.

Reviewable memory

The agent proposes. You approve. Only then does it count.

When the agent notices something worth remembering — a client's preference, a recurring meeting, a deadline — it proposes a fact and parks it. The fact never reaches a prompt until you have approved it. Rejecting it teaches the agent nothing.

3 facts pending

review
  • You run a design studio in Bologna.

    from "I run a design studio in Bologna"

    ApproveDismiss
  • You work Monday to Friday, 9 to 5.

    from "I work 9 to 5"

    ApproveDismiss
  • Marco Esposito is a recurring client.

    from email thread 18f2…9c

    ApproveDismiss

None of these are in a prompt yet.

Lives where you already chat

Telegram today. WhatsApp when Meta says yes.

Pair the agent to your phone in a minute. Pairing codes are single-use and short-lived, so a leaked screenshot is a leaked screenshot rather than a leaked account.

Telegram

live

Text it, or send a voice note and it will transcribe it. Approvals arrive here, so you can say yes from wherever you are.

WhatsApp

not yet

The adapter is written and tested, and signature-verified end to end. It waits on Meta’s business verification, which is paperwork rather than code.

  1. 01

    Open the bot

    Tap the link, or search the handle

  2. 02

    Send the code

    Single-use, and short-lived

  3. 03

    You are paired

    Chat normally from then on

Ready when you are

Quiet by default. Yours by design.