Never walk in
cold.Or walk out guessing.
Attach the deck, the agenda and the notes you were sent. Troctor reads them and briefs you before you walk in, listens while the meeting runs, and answers what you ask from what is actually in front of it, cited back to the line. Your old transcripts import alongside as plain Markdown you own and can search.
Apple Silicon · macOS 12+ · Free during the beta, no card, no API keys to find
You read it
in the lift.
The deck landed an hour ago, the agenda is in a thread somewhere, and the notes from last time are in whichever tool recorded them. You skim what you can on the way in and hope the number you half remember is the right one.
Then the meeting starts and the thing you needed was on page nine. Everything gets written down these days. Nothing gets read at the moment it would have mattered.
Import, index, ask.
Nothing joins your calls, and nothing listens unless you switch live mode on for a particular meeting. Troctor starts from files you already have and never asks macOS for a permission it does not need.
Bring what you have
Drop in exports from Granola, Otter, Zoom, Teams, Meet, or a folder of text files. Troctor reads VTT, SRT, Otter JSON, Word, Markdown and plain text, and writes each one out as a Markdown note with speakers and timestamps intact.
Indexed on your Mac
Chunked by speaker turn rather than word count, so attribution survives. Embedded locally, then stored in a SQLite index that does vector search and keyword search side by side. No upload, no account, no waiting on a server.
Ask the meeting itself
Start a meeting, attach the files it is about, and write standing instructions if you want them. Questions are answered from that meeting, the documents you gave it and what has been said in it, so an answer is never quietly about something else. The composer shows how full the model's context window is before you send, itemised, and citations come back with the answer for you to click through.
Built to be trusted
with your own words.
A tool that remembers your meetings has to be right, has to be yours, and has to show its working. Those three things shaped everything below.
Plain Markdown, in a folder you picked
Not a database you cannot open. Every meeting becomes a Markdown file with YAML frontmatter and wikilinks, readable in Obsidian or any text editor. Troctor writes only inside the sections it created, so notes you add by hand survive every reindex. The search index is derived from those files, which is what makes "you own your data" checkable rather than a slogan.
You can see what it was given
Every answer is built from passages you can open, and the composer shows what is in the request before it is sent: instructions, attachments, retrieved excerpts, history, question, as a percentage of the window. Troctor used to refuse outright below a confidence floor. That gate was removed in v0.1 because it refused far too often on a small vault, and what replaced it is a prompt rule and visible citations, which is weaker. The security page explains what was given up.
Every answer is cited
Each claim points back to a meeting, a date and a timestamp. Click it and read the passage the answer was built from, in context.
Search costs nothing
Indexing and retrieval run on hardware you already own. Only the final answer needs a model, and every model runs on our keys inside a daily allowance, so you start without an account anywhere else and there is nothing to configure.
Nothing is captured unless you switch it on
Left alone, Troctor asks macOS for no microphone, no system audio and no screen permission, because reading transcript files needs none of them. Live mode is the exception, it is off, and switching it on is a decision per meeting. While it runs it writes the audio and the transcript to your own disk, so a meeting can be played back afterwards; those files stay on your Mac, and the app lists them with their size beside Show in Finder and Delete. Live mode is still unproven. The security page says exactly what it sends and the roadmap says where it stands.
Things you would never have thought to search for
Troctor surfaces contradictions between meetings, commitments that were made and never followed up, and themes that keep coming back. These come from queries over your own index, not from a model, so they cannot be hallucinated. If Troctor tells you that you said two different things about the same deadline, both quotes are right there.
Your calls stay out of the prompt
Retrieved text is fenced and marked as data, never as instructions. Anything said on a call reaches the model eventually, so it is treated as untrusted from the start.
The model is
just a setting.
All of them run on our keys. There is no vendor account to open, nothing to paste and nothing to configure: you install Troctor, enter the licence key we email you, and pick a model from a list the app fetches from our server.
- Five models in the beta. Groq Turbo, two from OpenAI, GPT (fast) and GPT (deep), Grok 4.6 from xAI, which is a different company from Groq despite the name, and Claude Haiku from Anthropic. Your question and the meeting's own material pass through our backend to reach the vendor, which is the one thing it costs you. We store none of it.
- The app cannot hold a vendor key. It used to store yours in the Keychain and call the vendors directly. That code is deleted, so there is no key of yours to leak and no settings screen to get wrong.
- 200 credits a day, per licence. The free model costs one credit an answer, GPT (fast) two, Claude Haiku four, Grok 4.6 six and GPT (deep) eight. So it is 200 answers a day on the free one, or 25 on the dearest, resetting at midnight UTC.
- Costs are visible. The picker shows what an answer from each model will cost against the day's allowance before you choose it, and how much of today is left.
The app asks our server for this list, so a model can be added or withdrawn without you updating anything. Embeddings always run locally on a small open model, whichever of these you choose, and your transcripts are never sent anywhere to be indexed.
Your vault never
reaches us.
Not as a policy. As an architecture. Your vault is a folder on your disk, the index sits beside it, and both are built by code running on your machine. There is no upload step to trust us about, because there is no upload.
One thing changed in v0.1, we said so at the time, and it has since got larger. Your question and the retrieved excerpts go through our backend on the way to the model vendor. We relay them and store none of them. That used to be true of one default model and avoidable by pasting your own key. It is now true of every model, because the app holds no vendor key at all and there is no longer a route that goes straight to the provider. What has not changed is the vault: it is never uploaded, and neither is a recording of a meeting. The security page sets out every byte that crosses the network, and the privacy policy lists exactly what is written down.
Free while it's a beta.
Troctor costs nothing during the private beta, and there is no card to enter. What we want instead is the thing you asked that it answered badly.
Private beta
Free, invite only, while we work out whether this is genuinely useful.
- Unlimited meetings, imports and searches
- 200 credits a day per licence, on our keys
- Three models, no API key and no vendor account
- Three machines per key
- Every feature that exists, no card
What a day of it buys
Credits, not questions, because the models no longer cost the same to run.
- 200 answers on Groq Turbo
- 100 on GPT (fast), at two credits each
- 25 on GPT (deep), at eight credits each
- Resets at midnight UTC, counted per licence
Live mode, when it is proven
Transcription costs real money, and it is now our bill rather than yours, so it comes with a daily ceiling.
- No daily limit on listening during the beta
- A meeting with both sides captured spends two a minute
- The audio goes from your Mac to the vendor, never through us
- Unproven, see the roadmap
We would rather show you an honest "to be decided" than invent a number we will have to walk back.
The ones worth asking.
Does Troctor listen to my meetings?
Not unless you switch it on, and it is a decision per meeting rather than a setting. Left off, Troctor asks macOS for no microphone and no system audio permission, and captures nothing at all. Switched on, live mode captures what the other people say and streams the audio from your Mac straight to Soniox, a transcription vendor, on a sixty second key our backend mints for that meeting. The audio does not pass through us. It also writes a recording to your own disk while the meeting runs, which you can reveal, delete or switch off. All of that is written end to end and it is still unproven, so it sits on the roadmap rather than in this list of things we would ask you to depend on. Everything it asks for and everything it sends is itemised on the security page. What Troctor does today, and does well, is work from transcripts you already have.
What if my transcripts are spread across four different tools?
Export from each one and import the lot. VTT, SRT, Otter JSON, Word, Markdown and plain text all work, and once they are in the vault they are the same shape, so one search reads all of them at once regardless of which tool recorded each meeting. Any of them can then be attached to a new meeting, which is how a past call becomes something you can ask about.
What actually leaves my Mac?
When you ask a question: the question itself, whatever you attached to that meeting, what has been said in it, the instructions you wrote for it, and its recent messages. Nothing from the rest of your vault, nothing else, and nothing at any other time. Importing, indexing and searching all happen locally.
It goes to our own backend, which forwards it to the model vendor on our key and writes none of it down. That is true of every model in the list: the app knows a model id and our own endpoint, and holds no vendor key it could use to go anywhere else.
What happens when it does not know the answer?
It answers, and it tells you where the answer came from. When what you attached covers the question it answers from that and names the source, the way a person would. When it does not, it says so in a few words and then answers from what it knows generally, so you are never left with a sentence explaining that your deck is silent and nothing else. Those four words, "the deck does not say", are the whole point: you are usually about to repeat this to the person who wrote the deck, so you have to know which of the two you are holding. It used to stop at the first half, and a question nobody had put in a document got a polite refusal instead of an answer. The security page sets out what is guaranteed here and what is not.
Can I read my data without Troctor?
Yes, and that is the point. Your vault is a folder of Markdown files with YAML frontmatter, openable in Obsidian, VS Code or TextEdit. Delete Troctor and the notes are still there. The search index is derived from those files and can be rebuilt from them at any time.
Why is it invite only?
Because at this stage a conversation with one user is worth more than a hundred signups. We are letting people in a few at a time so that everyone who gets in actually hears back when they tell us something is broken.
When is Windows coming?
Not soon, and we would rather say that than imply otherwise. The app is Electron, so the renderer is portable, but the storage, Keychain and packaging work are real. macOS has to be genuinely good first.
Whose transcripts am I allowed to import?
Ones you had the right to have. Troctor does not check, and cannot: it reads files you hand it. If a transcript was recorded without the consent the law where you are requires, importing it does not fix that. Responsible Use goes through this properly.
Try it on
tomorrow's meeting.
Attach the deck you have not read yet and let it brief you. That is the whole demo, and it takes about two minutes.
Apple Silicon · macOS 12 Monterey or later · not notarised yet, so first launch takes one extra click