What the Gate can do
- “Reply to Novák that I’ll send the contract tomorrow.”
- “How many open invoices are overdue?”
- “Adjust this XML for Tatra banka.”
The Gate works out the intent. When unsure, it shows an “I understand this as…” card that you correct with one sentence. The answer is written live, with citations to your emails, documents and vault.
An answer with citations
A question about a contract, answered with references to the document and the email it comes from.
A chart on request
“Show a chart of invoices by month” and the Gate draws it, together with the table.
Real screens with sample data. The interface is in Slovak.
-
Conversations with history
Follow-up questions, picking up an old thread, search across history. Attachments are read locally.
-
Projects
Standing instructions, keywords, up to 20 files and the project’s tasks with a board. Every question in the project sees what matters from them.
-
Memory
“Remember that…”, “forget…”, “what do you know about me?” Overnight it proposes facts from conversations and from emails you sent yourself. Only what you approve is saved, to a readable file in your vault. The Scribe gets the facts for drafts too.
-
Procedures
Your own playbooks written as sentences, plus ready-made ones (Month close, Morning, Before a meeting). Start one by name or on a schedule and it runs step by step with confirmations.
-
Voice
It listens, transcribes, answers aloud and keeps listening, in the car too. Transcription and voice are both local.
-
Research
“Research on the web…” through your own SearXNG, with verified citations and a report in your vault. It can follow a topic weekly.
-
Artifacts
The answer as Word, Excel or Markdown in one click.
-
Thinking
It thinks longer on hard questions. Auto, Fast and Precise modes.
-
Charts
“Show a chart of invoices by month” and the thread gets a picture drawn from your numbers, not by the model, with the table under it. Charts are in the weekly summary and End of day too.
-
WhatsApp
Forward a message, a PDF or a voice note to your own “You” chat and Studio treats it as an instruction for the Gate. The answer comes back into the chat. Every other chat is dropped before it is even decrypted.
An intent you have confirmed eight times is decided without the model, from your own past choices.
Coding agent and self-learning
AI Development
The assistant has its own coding agent. Given a project, it makes an isolated copy of the repository, analyses the code and proposes or makes a change on a branch, with live progress and before-and-after screenshots of the site.
Every finding goes through a second pass that verifies it in the repository itself. Push, merge and deploy stay with you. It runs on local models under a separate account with no network.
It learns, but not on a free rein
-
Teacher
Every night it learns from your re-filings, edited drafts and invoice corrections. Each lesson carries evidence and is first replayed against 90 days of your decisions.
-
Scout
It looks for patterns in how you use the assistant and proposes features. It tests a new model overnight on your confirmed data. It proposes the switch; it doesn’t make it.
-
GEPA
Once a month it writes an addendum to the sorting prompt from your corrections and measures it on a held-out test. You accept or reject with one click.
-
Quality
A page, not a feeling: correction rate, first-try understanding in the Gate, drafts sent unedited, paid after the due date, days at zero.