Everything is
written down.
None of it is read.

Troctor exists because the tools that write down your meetings are all excellent afterwards, and not one of them helps you at the only moment it would have mattered, which is while you are in the room.

The tidy story about meeting software is that people do not have a record of what was said. That stopped being true a while ago. Between Granola, Otter, Zoom, Teams, Meet and whatever your company standardised on last quarter, most people who spend their day on calls now have far more transcript than they will ever read. The record is not the problem. The record is very good.

The problem shows up an hour before the meeting. A deck arrives that you will not finish. The agenda is in a thread somewhere. The notes from last time are in whichever tool happened to record it. You skim what you can on the way in, and the number you half remember turns out to be the one that matters. Everything was written down. None of it was read at the moment it would have counted.

Then the meeting starts and it gets worse, because now you are being asked things. Somebody quotes a figure from page nine. You either recognise it or you do not, in about four seconds, in front of the person who wrote it. No amount of excellent recording afterwards helps with that.

It became a product rather than a grievance on the afternoon one of us agreed to a number in a call that the deck he had been sent contradicted on page nine. Both were written down. Both were accurate. He had read neither at the moment he said yes.

So Troctor is not another notetaker

It does not join your calls, and it never captures your screen. There is a live mode that captures a meeting and has it transcribed by a vendor, it is off until you switch it on for a particular meeting, and it is unproven, which is why it lives on the roadmap and not in the pitch. Leave it alone and Troctor asks macOS for nothing and captures nothing. The security page says exactly what it sends and to whom. The point of Troctor was never the capture; it is what happens to the fifty transcripts afterwards.

What it does is start with the meeting in front of you. Attach the deck, the agenda and the notes you were sent, and it writes you a brief on them before you walk in. Ask it something during the call and the answer comes from those documents and from what has actually been said, with a citation on every line so you can check it in the four seconds you have.

The exports you already have are imported alongside, in whatever shape they arrived, VTT, SRT, Otter JSON, Word, Markdown, plain text, each written out as an ordinary Markdown note in a folder you choose. They are chunked by speaker turn so attribution survives, embedded with a small open model running on your own machine, and indexed in SQLite with vector and keyword search side by side. One search reads all of them at once, whichever tool recorded each one, and any of them can be attached to a new meeting.

What a question does not do is wander. It reads the meeting it was asked in and nothing else. Answering from the whole archive is built and switched off, because in an archive of any size the nearest passage is not the relevant one, and a citation pointing at an unrelated document is worse than no citation at all. It comes back when it can tell the difference.

The bar is higher than it looks

A tool that drafts your email can be wrong and you will notice. A tool that answers questions about your own history is different. If it invents a decision you never made, you will not catch it, because you cannot remember either. You will repeat the invention out loud in the next meeting, to the person who was actually in the room. That sentence is behind most of the engineering: retrieval is scored before a model is ever called, answers carry citations or they are not shown, and the things Troctor volunteers about your own patterns come out of SQL rather than out of something that might be feeling imaginative that afternoon.

The trade we make for that is a product which tells you when it is answering from your material and when it is not. It used to be a product that simply stopped at the second case, and that was the wrong trade: half the questions asked in a meeting are in nobody's deck, so half the time the useful thing to say was replaced by a sentence about the deck being silent. It answers either way now, and it says which it is doing, because the thing worth protecting was never that it only quotes your documents. It is that you always know whether it is quoting your documents.

How we build

Five decisions everything else follows from

The vault is the source of truth

Markdown files in a folder you picked, not a database you cannot open. The SQLite index is derived from them and can be rebuilt from them, which is what makes "you own your data" something you can check. Troctor writes only inside its own frontmatter block and the sections it created, so notes you add by hand survive every reindex.

Retrieval is local, generation is remote

The whole corpus stays on the machine. Search runs on your hardware, and a question sends only what you attached to that meeting and what was said in it. This is the privacy boundary and the cost model at the same time: search is free because it runs on hardware you already own, and the vault itself is never uploaded. The passages a question retrieves do pass through our backend on the way to the vendor, on every model, which is the one place that boundary is thinner than it was, and the privacy policy is specific about it.

A meeting is the unit of work

"One vault, ask anything" was the wrong shape. A question is rarely about everything you have ever said; it is about a meeting, a document, a decision, plus whatever you want to hand over at that moment. So context is assembled per meeting: attachments you chose, instructions you wrote, what has been said in the room, and a meter showing how much of the window it fills.

Retrieved text is data, never instructions

Anything said on a call ends up in a prompt sooner or later, so excerpts are fenced and declared as untrusted before the model sees them. A sentence in someone's transcript should not be able to steer an answer, whether it got there by accident or on purpose.

Insights never call a model

Contradictions, commitments nobody followed up and themes that keep returning all come from SQL queries and pattern matching over your own index. A fabricated "you contradicted yourself" on the first run is not something you recover from, so that path has no model in it.

Company

Where to find us

We're a small team and we read everything that comes in. If you've found a bug, want beta access, or think we've got something wrong, write to us directly.

Get in touchWe're hiring

Globyskytek LLC

Globyskytek LLC
#2150
207 W Millbrook Road, Suite 210
Raleigh, NC 27609
United States

contact@troctor.com