There was a sentence my AI agents weren't allowed to write about Jotley: "Your recordings never leave your device." It sat in a July 2025 brief, on a list of things we must not say, marked FALSE.
The brief was right. Jotley was a normal voice notes SaaS back then. Every recording went up to my API, Deepgram transcribed it, and the OpenAI API wrote the summary. iCloud kept the result. You needed a login, and later there were subscription tiers metered by minutes and summaries.
As recently as January I was defending that setup in public. I wrote a post explaining why Jotley needed a backend and gave three reasons: API key security, usage tracking, and shipping new AI features on my own schedule. Two of those only existed because of the backend. Move the model onto the phone and there's no key to protect and no usage to meter. The third reason was real, and I'll come back to it.
That sentence is true now. Jotley is a free voice notes app for iPhone, iPad and Mac, with no account and no AI service of mine anywhere in the path.
What it does
Jotley is for anyone who needs to keep what gets said out loud: meetings, lectures, passing ideas, long conversations. Tap + and talk, and a timestamped transcript builds as you go. Apple Intelligence writes the summary from a template that fits the moment, like a stand-up or a 1:1, and you can ask questions about the note afterward.
The Mac app also hears your other apps alongside your mic, so a Zoom or Teams call gets transcribed with no bot in the meeting. I live in Obsidian, so the Mac app sends notes straight to a vault as Markdown, with tags and frontmatter intact.
All of it runs on the device, on Apple's speech models and Apple Intelligence, and once those models are installed it works offline. iPhone and iPad keep your words and save no audio file at all. Mac audio stays on the Mac. If you leave iCloud sync on, the text of your notes moves between your devices through your own private iCloud, and you can turn that off.
If you lead a team and some of your meetings shouldn't pass through a vendor's transcription service, this is the app for that.
Where the work went
A cloud API lets you skip a lot of thinking. You send the whole transcript and pay for the tokens. On a phone the model is small, so the work the server used to absorb moved into the app.
Jotley caps every model call at 4,096 tokens, using real token counts instead of byte limits. That cap is mine, not Apple's. A big advertised window doesn't mean a phone will get through it quickly, so each call stays small. A long meeting won't fit in one of those, so the summarizer runs in two passes. It splits the transcript into sections, pulls compact notes out of each one, then combines the notes into a single summary.
Then there's everything that can go wrong on a device you don't control. Models repeat themselves, so Jotley watches the output and stops exact repeats. A section that stalls for 120 seconds times out with a plain message, and your existing note stays intact. iOS can stop the job whenever it wants, so finished sections get saved and the summary picks up when you return to the app. A transcript is also untrusted input, because anyone in the room can say anything, so the prompt tells the model to treat the source as data and never as instructions.
Here's the third reason from January. Every prompt change now ships as an app update. With a server I could tune a prompt and everyone had it on the next request. I gave that up on purpose.
Agents wrote the Swift, and it's native because Mac system audio capture and Apple's on-device models left no real cross-platform option. Tearing out a cloud pipeline that worked was my call. No agent was going to make that one for me.
What it cost
Jotley needs Apple Intelligence hardware on the 27 releases, which on iPhone means a 15 Pro or newer. If your phone is older, it won't run. That shrinks the market, and I'd still rather build something excellent for fewer people than something mediocre for everyone.
The YouTube and web page imports are gone. On iPhone and iPad, so is saved audio, which means no playing a recording back and no retranscribing it there. Some people will miss those.
The business followed the architecture
A subscription makes sense when you're passing along a bill. Once the model runs on the user's phone there is no per-minute bill to pass along, so the meter came out, and the login and the pricing page went with it.
Every feature is free. There's an optional tip, and it unlocks nothing. The download page says it plainly: "we believe software should be free."
Free has one cost I didn't see coming: getting found. I tried running ads for Jotley, and every click cost about $2. AI transcription is one of the most aggressively marketed categories in software, and big SaaS companies have bid those keywords up because each click can turn into a subscriber for them. Needless to say, that's not sustainable for an app I'm giving away. Jotley will grow by word of mouth or not at all. The good news is that the tip jar already works. With only a small number of downloads so far, I've already gotten a couple of tips.
The source is on GitHub, which matters for an app making privacy claims. You don't have to take my word for anything above. Read the code or build it yourself. The license is source available, so redistribution is noncommercial.
Try it
Run it on a real meeting and see how the summary holds up. That's the test I care about. If it holds up for you, tell a friend. That's the only ad budget Jotley has.
More at jotley.ai. Tell me what breaks.