I never said the horse should be consul.
It is in the acta diurna, Caesar.
With your name on it.
Your agent knows what was said in the meeting.
Acta writes down your Google Meet call with every participant’s name and hands it to Claude Code or Codex. The audio never leaves your Mac.
01 / 05
Caesar
I never said the horse should be consul.
Mark Antony
It is in the acta diurna, Caesar.
Octavian
With your name on it.
New context
transcript.md3 speakers · 00:32:14
Your agent already knows
The Forum
- Client contact
- Marcella
- Previous decision
- Send a draft for internal review before making any offer.
- Project context
- Already with your agent
transcript.md added to context
You
Turn today’s decision into an internal review checklist. Flag what still needs confirmation.
Agent
I connected this to The Forum and found two points to verify before acting.
- Does Friday mean internal review or sending it to the client?
- Confirm the recipient’s name is Marcella.
Acta is the only Google Meet tool that transcribes Czech and Slovak entirely on your Mac and names every speaker automatically.
- Transcribed on your Mac
- whisper.cpp runs on the Mac itself. No audio leaves the computer.
- Real names, no guessing
- Google Meet supplies each participant’s name, and Acta assigns it to every sentence automatically.
- Czech and Slovak first
- The app speaks Czech, Slovak and English.
- No bot in the call
- A Chrome extension and a Mac app. Nobody joins the meeting.
- Your agent takes it from there
- The transcript goes to your own AI agent, such as Claude Code or Codex.
What the other ways of doing it leave out
- Cloud tools
- Your audio goes to their servers.
- Google Meet’s own transcripts
- Do not support Czech or Slovak. Google Meet help
- Local tools that label “Speaker 1” and “Speaker 2”
- Nobody tells you who is who. You work it out afterwards.
- Acta Diurna
- Czech and Slovak, on your Mac, every sentence under a real name.
Your agent already has the project.
Now it gets the transcript.
Speech recognition will miss words. Acta keeps who said what and when. Your agent can work with that record alongside the names, documents and decisions it already knows.
You had the meeting. Let your agent read the transcript.
- Claude
- Codex
- Another AI agent
Let your agent help with setup Draft guide · macOS and Windows instructions to follow
transcript.mdFrom Acta · Original
“Send the forum proposal to Marcellus on Friday. Use the terms we agreed last time.”Antony · 00:18:42
An illustrative transcription with words that need checking.
- Project
- The Forum
- Client contact
- Marcella
- Previous decision
- Send a draft for internal review before making any offer.
A possible
next step.
Your agent With project contextWithout project context
- Connect to the project
- This appears to concern the Forum proposal for Marcella.
- Draft the follow-up
- Prepare the proposal for internal review, using the previously agreed terms.
- Check before acting
- Does “Friday” mean the internal review or sending it to the client? Confirm the recipient's name.
- What the record says
- A proposal is due on Friday.
- Missing context
- Which project? Who is “Marcellus”? What terms were agreed?
- Check before acting
- Ask for the project, recipient and previous decision before drafting the follow-up.
Illustrative workflow, not a live AI response. Acta writes the record; your own agent handles the follow-up. Context helps interpretation, but does not guarantee a correction. The original transcript stays intact.
The model and settings are already chosen.
Most local transcription tools give you parts to assemble and tune. Acta Diurna comes preset for live Google Meet calls, including Czech, Slovak and English in the same conversation. acta setup downloads and verifies the chosen Whisper and VAD models once. After that, acta live writes the named transcript while you meet.
acta live· Q3 Roadmap Sync· viewing (PID 68438)
capturing 4 participants + mic
queue · captured 14m30s · transcribed 14m18s · backlog 12s
CHANNELS
Nadia Kerr2m 50s
Tomas Bergqvist2m 50s
Priya Raghunathan2m 50s
Marcus Adeyemi2m 50s
You (mic)2m 50s
mic state unmuted
transcript
## [00:14:12] Nadia Kerr
One caveat: the staging cluster still runs the old schema.
## [00:14:15] You
Fine by me. I'll send the summary round once this call ends.
## [00:14:18] Tomas Bergqvist
Let's give it two sprints and look at the numbers again.
## [00:14:21] Priya Raghunathan
So the pilot rollout slips a week, but only for the Berlin team.
## [00:14:24] Marcus Adeyemi
Can we get the migration numbers before Thursday's board call?
## [00:14:27] Nadia Kerr
I'll take that one. The export job is already half written.
## [00:14:30] Tomas Bergqvist
One last question on the pricing page before we wrap up.
## [00:14:33] Priya Raghunathan
Keep the current price for the pilot. We can review it after two sprints.
4 participants + mic · 8 segments · 5 names
microphone: MacBook Pro Microphone
8 windows transcribed
microphone: MacBook Pro Microphone
8 windows transcribed
Open it in your browser, or from the terminal.
The record is one folder on your disk. acta gives you two doors to it: a panel served by your own machine that the extension opens beside the call, and the command line you already live in. Nothing here reaches a server.
recording
3 speaker channels + your mic
- channels
- 3
- mic
- on — MacBook Pro Microphone
- last frame
- now
- run
- meet-1788784283
Octavian
Minute it as a proposal.
Local records
Stored on this machine. Changes stay local.
3 runs
| run | meeting | size | Open |
|---|---|---|---|
| meet-1788784283 | Consulship handover | 1.7 GiB | |
| meet-1788511608 | Grain contract review | 851.6 MiB | |
| meet-1788508868 | Senate stand-up | 687.2 MiB |
Showing Finished · transcript.md
## [00:00:02] Caesar
I intend to make my horse a consul. The paperwork is the only obstacle.
## [00:00:11] Mark Antony
The horse has no opinions on the grain price, which puts it ahead of the senate.
## [00:00:19] Octavian
Minute it as a proposal. We can lose it in committee like everything else.
Canonical transcript is absent. This run does not contain transcript.canonical.md.
$ acta ls
run date time dur meeting windows size
$ acta show meet-1788784283
## [00:00:02] Caesar
I intend to make my horse a consul. The paperwork is the only obstacle.
## [00:00:11] Mark Antony
The horse has no opinions on the grain price, which puts it ahead of the senate.
## [00:00:19] Octavian
Minute it as a proposal. We can lose it in committee like everything else.
The meeting above is invented; the three surfaces are the real ones acta ships.
From the call to your folder,
and nowhere else.
Three steps, and none of them leave the room you are sitting in.
- Google MeetAvailable now
- Microsoft TeamsSupport in development
- ZoomSupport in development
In the call
Nobody joins. No bot, no waiting room, no fourth participant called Notetaker — and because Meet already knows who is speaking, the name arrives with the voice.
On your machine
The sound crosses from the tab to Acta Diurna, and stops there.
In your
folder.Transcribed on your own hardware, into a document you can open, move or delete.
See what it writes
Never leaves this machine
The audioNever
The transcriptNever
Who was in the roomNever
No account, no tracking, and no company holding anything of yours. The only thing Acta Diurna ever sends is a version number, when it looks for a new release. Delete the folder and the meeting is gone.
Acta Diurna learns your words without learning your voice.
Every practice has words no transcription has met before: client names, matter numbers, the way your firm writes a statute. The usual answer is to train a model on your recordings. We tried that and refused it — the words were not being misheard, they were being written as ordinary words, and training cannot fix a spelling convention.
You can write the list yourself, or ask an agent to draft it for review. It is applied afterwards. Every change is recorded with the rule behind it, and the original transcript is never altered, so a wrong rule is easy to see and easy to undo — which is not something a trained model can offer. One shared list keeps the firm’s vocabulary in one place instead of in eight people’s heads.
Your agent can help write the dictionary.
Claude / Codex + your project vocabulary
From the transcriptKnown project term
“cross undertaking”cross-undertaking in damages“harlow bee cee”Harlow BC“matter four one seven”Matter 417Three suggested rules. Ready for your review.
[[rule]]
spoken = "cross undertaking"
canonical = "cross-undertaking in damages"[[rule]]
spoken = "harlow bee cee"
canonical = "Harlow BC"[[rule]]
spoken = "matter four one seven"
canonical = "Matter 417"Illustrative workflow: give your agent the transcript and project vocabulary, then review its suggestions. This demo sends and saves nothing. The original transcript stays unchanged.
What it needs, and what it is not.
- Platform
- macOS on Apple Silicon. Windows support is planned. Linux is not supported.
- What you install
- Two things, and both are needed: the Acta Diurna app on your Mac, and an unlisted Chrome extension you install from a link. The extension on its own does nothing.
- Calls
- Google Meet, in Chrome, with the extension loaded. Support for Microsoft Teams and Zoom is in development and is not available yet.
- Consent
- Acta Diurna is your recorder, not a participant: it never joins the call and never announces itself, so telling the room is a decision you make, the way you would with a notepad. Most places require it.
- Retention
- Nothing is deleted for you. A meeting stays on your disk until you delete the folder, and no copy exists anywhere else.
- Not a summariser
- No AI summary, no action-item extraction, no chat over your meetings — on purpose. It writes the record and hands it to whatever agent you already use.
- Not a service
- There is nothing to sign into. If we disappear, the program on your disk keeps working.
Buy a licence. That is the whole relationship.
No account to create, nothing to log into, no service in the middle. It works with the internet switched off, and an expired licence never stops a recording in progress or takes away a transcript you already have.
One seat, yearly€69per year
- Everything the tool does
- Updates for the year
- Your recordings stay readable either way
Five seats or more€49per seat, per year
- The same tool, less per person
- One invoice, one VAT number
What a seat isA seat is one person, on up to two of their own computers, such as a laptop and a desk machine. A licence is not shared between people: a team buys one seat per person.
Founder priceThe first 50 customers keep their price for as long as they are held; after that the list price goes up. This is the only discount there will be — no seasonal sales, no coupon worth waiting for.
Compare your options
Acta DiurnaCloud notetakersLocal open source
Where the audio goesYour machineTheir serversYour machine
Who joins the callNobodySome send a botNobody
How it knows who spokeRead from Google MeetGuessed from the audioGuessed from the audio
What you end up withA file in your folderAn account you log intoA file in your folder
Let your agent
set up Acta.
Install the Acta Diurna skill, then ask your agent to set up Acta. The skill holds the installation steps, and afterwards it shows the agent how to work with your meetings.
The skill is current: it names the published release, its checksum, and which parts have never run on real hardware.
1. Install the skill
npx skills add gaboe/homebrew-acta --skill acta-diurnaRun it in a terminal; it needs Node.js. It works with Claude Code, Codex and other agents. Listed on skills.sh.
macOS
Apple Silicon
- Installer
brew tap gaboe/actathenbrew install acta- Setup instructions
- Published and installed from the tap. brew trust gaboe/acta has to come first, then acta setup downloads the model once.
Windows
x64
- Installer
scoop install https://raw.githubusercontent.com/gaboe/homebrew-acta/main/scoop/acta.json- Setup instructions
- Published, and never run on a Windows machine. It also needs whisper-cli.exe on PATH from whisper.cpp. Whoever installs it first is asked to send back what acta report prints.
One skill. A link your agent can read.
The skill is in a public repository, listed on skills.sh and served from this website. The same repository holds the formula, the Scoop manifest and every release archive with its checksum.
- GitHub
- gaboe/homebrew-acta — the skill, the formula, the Scoop manifest and the release archives. The source repository is private during the beta.
- skills.sh
- gaboe/homebrew-acta/acta-diurna
- Browser extension
- Chrome Web Store — an unlisted listing, reachable by this link. Chrome keeps it updated.