Azynote 1.3.2: Bring Your Own Agent CLI
Most of you are already paying for an AI subscription. Probably more than one.
There is a decent chance you have an agent sitting in your terminal right now that is smarter than anything Azynote would call on your behalf, and that you are paying for twice: once for the terminal agent, once for whatever the app is using.
This release is about stopping that.
Your terminal agent writes your OnePagers
If you have Claude Code, Codex, opencode, Cortex, or anything else that takes a prompt and answers on the command line, Azynote can hand its thinking to it. You type the command once in Settings > AI > CLI, and that is the whole setup.
No API key to paste. No second bill. Your existing subscription, your model choice, your machine.
This existed before 1.3.2, and it half worked. Four of the six places Azynote thinks used your command. The other two, OnePagers and Smart Actions, quietly ignored it. If you had set this up with anything other than Claude Code, those two were failing while everything else looked fine.
Now that field drives all of it. Including which model gets used, which you set in the same command where you can see it.
Check it works before a real meeting depends on it
Next to the field there is a Test command button.
It runs your command once, with a throwaway prompt, and shows you the command it actually spawned next to whatever came back. About a second. Nothing saved, nothing started.
This sounds minor and it is the thing I would use first. A command line has a lot of places to put a typo, and a stray character in the middle of a long one is genuinely invisible until you see the pieces laid out separately. Better to find it here than on the Friday afternoon when you have six sessions to summarise.
If it goes wrong, you get told
The reason the test button matters is that the alternative used to be worse than an error.
A CLI can run, finish without complaining, and hand back something Azynote cannot use. When that happened you got a OnePager with a title on it, nothing underneath, and a green checkmark saying it was done.
Now that fails, and it fails out loud, with whatever your CLI actually said in the message. A OnePager that cannot be built is a message on your screen, not a document you discover is empty three days later when you go looking for a decision.
Your memory back, one model at a time
If you run Azynote on the local model, the indicator at the top of the window shows you what is loaded.
It has always listed two: the chat model, which is the large one, and the search model, which is small and powers indexing. But the only thing you could do was release both at once.
Now each line is a button. Hand back the big one and keep search running, which is usually what you want, because the search model comes back in about a second while the chat model takes half a minute.
The setting is now called Keep the chat model ready, because the chat model is all it ever kept.
And the memory figure is now correct. It was reporting the model's own size while the process around it could grow considerably larger. If you are on a 16 GB Mac and you have been deciding whether you can afford to keep the model loaded, you were deciding on a number that was too small. The one you see now is the one you get.
Also in this release
- Long meetings finish cleaning up. Transcript improvement on a long session could run out of time and fall back to the uncorrected version without saying so, which meant a OnePager built on raw text with unattributed quotes.
- New OnePagers turn up in search. Some were being written and never indexed, so they were invisible to chat and to semantic search.
- A provider you never set up is no longer red. Not configured is not the same as broken, and only a key that was checked and found invalid gets the error colour now.
- Improving a transcript by hand no longer causes a second pass when you then generate a OnePager.
Getting it
macOS, free, no account, nothing to activate. If you already have Azynote installed, replace it in Applications and your workspaces carry over untouched.
If you run your own CLI, tell me how it goes
This whole path exists because a handful of people asked for it, and it took someone pointing it at a CLI I had never tried for me to find out how much of it only looked like it was working.
So if you set this up, I would like to know which agent, and whether it held. Especially if the answer is that it did not.
Happy meetings.