What's built, and
what isn't.

Most roadmaps are a wish list with the dates removed. This one separates three things: what you get if you install Troctor today, what is actually being worked on, and what we decided against.

Last updated 19 August 2026

Shipping now

In the beta build.

Everything here works today. If any of it does not work for you, that is a bug and we want to hear about it.

Import what you already have
VTT, SRT, Otter JSON, Word, Markdown and plain text. Exports from Granola, Otter, Zoom, Teams and Meet all land as one of those.
A vault you own
Every meeting becomes a Markdown file with YAML frontmatter and wikilinks, in a folder you chose. Obsidian opens it. So does TextEdit.
Local indexing
Speaker-turn chunking, embeddings computed on your Mac, stored in a SQLite index that does vector and keyword search side by side.
Meetings
The left panel lists meetings. Start one empty or from an imported transcript, attach files, paste text, and write standing instructions that steer just that meeting.
Search across everything
One search reads every meeting you have imported, regardless of which tool it came from originally. Answers are a separate thing and read only the meeting they were asked in.
A brief before you walk in
Attach the deck and the notes, and the meeting opens with what they commit to, what matters in them and what they leave open, cited.
A context meter you can read
Before you send, the composer shows how full the model's window is and what is filling it: instructions, attachments, retrieved excerpts, history, question.
Citations
Every answer points back to the meeting, date and timestamp it came from. Open the passage and read it in context.
Answers with no API key
Every model runs on our keys, inside 200 credits per licence per day. An answer costs one credit on the two free models, two on GPT (fast) and eight on GPT (deep). The trade is that those requests pass through our backend on the way to the vendor, and there is no longer a setting that avoids it.
Insights
Contradictions between meetings, commitments that were never followed up, themes that keep returning. Derived from queries over your index, never from a model.
Your choice of model
Groq Turbo, and GPT (fast) and GPT (deep) on OpenAI. The app fetches the list from our server, so one can be added or withdrawn without you updating anything.
Removed in v0.1: the confidence floor
Retrieval used to refuse below a score, and it refused far too often on a small vault. The model is now always called. This was a loss as well as a fix, and the item below is what would let us bring the floor back properly.
Search runs on your Mac
Import, index and search need no network, once the search model has been fetched the first time, about 33 MB per Mac. Answers always need one: the option of a local model went with the provider layer.
Being worked on

Next, in rough order.

No dates. A date on this page from a one-person team is a guess dressed as a commitment, and you would be right to hold us to it.

Answer quality, measured rather than guessed

Troctor used to refuse to answer below a hand-picked confidence floor. That floor was removed in v0.1 because on a small vault it refused constantly, which is worse for a new user than a hedged answer that cites what little it found. The honest position is that we never knew whether the number was in the right place, because measuring it means testing against real private transcripts, which by design we cannot see. Building an evaluation harness that runs on your machine and reports only a score is the first thing on this list, and a measured floor is what it would buy back.

Live mode: built, and unproven

This is the feature people ask about first, and all of it now exists in the repository. It captures what the other people say, streams it to Soniox for transcription while the meeting runs, writes the transcript into that meeting, and occasionally puts a cited one-line suggestion on a small window over your call. The capture layer landed too: system audio through a Core Audio process tap, your own voice through the microphone, each one its own stream so the transcript can say who spoke without guessing.

This entry used to say it has never been run, and that has been withdrawn: real sessions have happened, audio has reached the vendor and come back as text, and the faults the first of them exposed have been fixed. A device change no longer ends a meeting, and a recording is written continuously to your own disk so a crash cannot take a call with it.

What has not happened is a week of somebody else's real calls. Tests prove the code does what it was written to do and prove nothing about what macOS, Chromium or Soniox do on a machine we have never seen, and the transcription path changed again when the vendor account moved to us. So it stays here rather than moving up to the list of things that work, and the honest description is built and unproven rather than built and finished.

Four things about it that will not change, and which are on the security page in full. It captures nothing unless you switch it on for a particular meeting. The audio goes from your Mac straight to Soniox and never through us, on our account and our bill, with no daily limit during the beta. It writes a recording to your own disk, at about 115 MB per channel hour, which you can reveal, delete or switch off. And it sends the audio of the meeting to a company that is not us.

Watching a folder

Right now you import by hand. If your existing tool already drops transcripts into a folder, Troctor should notice and index them without being asked. Small, useful, and the thing that turns this from a tool you remember to use into one that is simply current.

More sources, properly parsed

Every export format has its own idea of what a speaker label looks like. The parsers handle the common cases and fail politely on the rest. Each new source is a small amount of work and a large amount of testing, and we would rather add them in response to a file that broke than in anticipation.

Windows

Asked for, and not soon. The interface is Electron and would port. The storage layer, the credential handling and the packaging would all need doing again. macOS has to be genuinely good before that is a sensible use of the only pair of hands here.

Decided against

Things we're not doing.

Some of these were built into earlier plans and cut. Knowing what a product refuses to do tells you more than the feature list.

Uploading your vault
A server-side index would make search faster and sync trivial. It would also mean we hold every meeting you have ever had, and every privacy claim on this site would become a policy you have to trust rather than an architecture you can check. Not worth it.
Marking up model tokens
The obvious business model is to resell inference at a margin. It also means the incentive is for Troctor to send more tokens rather than fewer. The daily allowance is capped and every answer costs us money rather than earning it, which points the other way: retrieving eight passages instead of eighty is in our interest as well as yours. When the beta ends, the thing being sold will be the app rather than your usage.
Generating the insights
Asking a model to find contradictions in your history would produce a longer and more interesting list. It would also produce a list that is sometimes invented, and a fabricated "you contradicted yourself" on first run is not recoverable. They come from queries instead, so both quotes are always real.
Hiding from screen shares as a selling point
This entry used to say the mechanism was not in the product. It is: the overlay window that live mode uses asks to be excluded from screen capture, so a private note stays private on a shared screen. On macOS 15 and later that request no longer works, because Apple changed the compositor and the capture framework every meeting app now uses reads the finished image rather than the individual windows. We found out when a beta user shared their screen and the panel went with it. The app now says which of the two situations you are in, on its own face, instead of letting you discover it in front of other people, and it says the thing that does work: share a single window rather than the whole display and the panel is not in the picture on any version, because macOS captures a chosen window on its own. It is a switch rather than a constant, because demonstrating Troctor, or working beside somebody who is meant to see the panel, are real cases; it is hidden by default and hidden wherever the setting is missing or unreadable, and both windows say in words which state they are in. What has not changed is the position. It is there so your own notes are not broadcast to the room by accident, not so anyone can be misled about whether the tool is running, and we will not market it as a way to be undetectable. See the security page.
Depending on another app
There was a plan to expose the vault to editors and assistants over a plugin protocol. It made Troctor useful only to people already running something else. One app, one backend behind it, and nothing else for you to install or sign up to.
An account you need to use it
The licence key is checked once at activation and then the app works offline indefinitely. Even if we revoke a key, nothing on your Mac stops working. Your notes are your files.
How this list changes

Mostly by beta users telling us something is wrong. The order above reflects what the twenty or so people in the pilot actually run into, not what sounds best on a page. If you are using Troctor and something on this list is in the wrong place, say so.

Want to be one of them?

The beta is free and invite only. We let people in a few at a time so everyone who gets in actually hears back.

Ask for a key See how it works