These are the questions an agent asked me before writing this case study, and my answers exactly as I typed them, typos included. I answered them in a form, and the agent turned them into the article.
What the agent found
Before asking anything, the agent read the code, the commit history and the live app. This is the summary it showed me.
A native macOS 26 menu bar app. ⌥ Space opens a floating Liquid Glass panel over any app without stealing focus; answers stream in, and chats live only in memory, deleted 30 minutes after the last message. No backend, no analytics, API keys in the Keychain.
It grew three modes: LLM (Apple Intelligence on-device by default, Anthropic, OpenAI, Gemini, OpenRouter, Ollama/custom), Agent (Claude Code, Codex, OpenCode with MCP servers, Allow/Deny permission cards and its own present_files MCP tool), and Decision (System One models like TypeSafe’s Jev answering yes/no, choices or levels with confidence, plus Live and Improve). On top: context capture (selection, clipboard, screenshot), ⌘K actions, Show What Changed diffs, tear-off notes, presets, 8 word games, a usage/cost dashboard and a meraline:// URL scheme.
Swift 6 with strict concurrency, SwiftUI + AppKit, ~32k lines of app code and 825 tests. Signed, notarized DMG, Sparkle auto-updates every 15 minutes with a beta channel, Homebrew cask, build attestations — all set up on day one. README screenshots are generated by the test suite with ScreenCaptureKit.
The trailer is a 96-second Remotion film cut to a 116 BPM soundtrack assembled from GarageBand Apple Loops by a script, recorded by driving the real v1.4.0 app with scripted input. Made on the evening of day two. The rendered video isn’t in the repo and I couldn’t find it published anywhere.
Fact check
Claims the agent wasn’t sure about, and my verdict.
- Work started 2026-09-25; v1.0 shipped about an hour after the first commit (which was already ~2,700 lines). Last commit 2026-10-07.
Correct. - 249 commits over 11 active days, 15 releases in 13 days. 248 of 249 commits are co-authored by Claude models (Fable 5.1, Opus 5.5, Opus 4.8).
Correct. - Downloads so far: 25 DMG, 79 zip (Sparkle updates + Homebrew). 10 stars, 1 fork (renamed “intentful”), no issues or PRs.
Correct. - Decision mode uses TypeSafe’s Jev and other System One models via TypeSafe, OpenRouter, Ollama and Ollaya — added on 2026-09-29.
Correct. - The trailer was inspired by the promo on getdroppy.app and was shared on Discord (two compressed copies were rendered for it), possibly also YouTube/X.
Correct. - The trailer shows v1.4.0, so it predates Decision mode. A separate “Live decisions” promo clip was recorded on 2026-10-03.
Correct. - Meraline is MIT-licensed, free, and not on the Mac App Store (no sandbox, Developer ID signed).
Correct.
Origin & vision
What made you build Meraline?
A specific moment or annoyance? What were you using before (ChatGPT app, Raycast AI, Spotlight…) and what was missing?
Raycast made BYOK paid, and I was looking for alternatives for quick chat AI, but couldnt find any good and open source.
Where does the name “Meraline” come from?
Ephemeral -> Meral -> Meraline (female assistant)
Why is “let it go” the core idea instead of a privacy toggle?
30-minute chats, nothing on disk. Do you personally never want history? Has anyone asked for persistence?
Inspired from quick AI, chats are not meant to be persistant, you should just ask, get answer, and its gone. It remains for a bit in case you closed it by mistake, or need to recall something you just talked about.
Who is it for?
Yourself, developers, privacy-minded people, Appwrite colleagues…?
Anyone, but mainly me and my friends.
Building it
How did you build it with Claude?
248/249 commits co-authored by Claude. How did you split work between models, what did “several sessions edit this checkout at once” look like, and did the 96 KB AGENTS.md make a real difference?
Most of work with Opus, with some critical flows by Fable.
v1.0 shipped an hour after a 2,700-line first commit — was there a prototype before the repo?
No prototype, directly to 1.0, why not. product is best tested when used.
Why native Swift for macOS 26 only, and how was working with Liquid Glass?
Precisely for liquid glass, and best compatibility with macOS. no plan for other platforms right now.
What was hardest?
Streaming performance / rendering, Release pipeline (signing, notarization, Sparkle)
This project was my first experiene with Apple developer program, learning a lot of quirks of all the keys, signatures, certificates, …
Which providers do you actually use day to day, and why support so many?
I use OpenRouter for LLM, because I like ability to switch between any models, for a little extra price. Plus I prefer pay per usage for LLM, because I use it very very little..
For agent, I use Claude max, because I have company one that I can use. I would still use Claude probably, just cheeaper plan.
And for decision mode, I use Jev, as it’s early and there arent many options. I will be switching to OpenRouter when I run out of Jev credits.
Why build a full notarized DMG + Sparkle + beta channel + Homebrew pipeline on day one?
I wanted to learn to do this properly on this project. I also took a lot of inspiration from OpenClip to which I controbuted a few pull requests before startign Meraline project.
Screenshots are generated by the test suite. Did that pay off? Would you reuse it?
I love it. Automated screenshots are huge win for quality assesment.
Features & scope
What’s your connection to TypeSafe / Jev and Ollaya, and why add Decision mode?
Partnership, your own thing, early adopter? Who are Live and Improve for?
Decision mode was idea due to rising popularity on this topic, and I saw huge potential that no other AI clients did yet - both realtime text decisions, as well as bulk processing.
Why word games in an AI utility? Does anyone play them?
I play them. I like playing them when on boring meets to stay sharp. For example I like rhyme training, and word crafting, liek scrabble.
From “tiny app for quick questions” to three modes and 15 releases in 13 days — intentional? What would you cut?
Intentional, and I like it this way. Passion allowed me to go full force for 2 weeks, building a fully finished product that serves me and my friends great value. both me and agent are pretty-much out of ideas what could be improved next, only tiny details.
Your favourite feature or detail?
Realtime decision mode. I can add question like “Flirt?” and learn my flirting skills, hihi.
The trailer
Why make a beat-synced product film the day after launch, and why Remotion instead of a video editor?
Remotion was tool agent waas confident with for ability to make draft, and then keep making edits to improve quality with my experience.
What was the hardest or most fun part of making it?
Driving the real app with scripted input, the Apple Loops soundtrack script, keeping the edit in sync after re-recordings…
I liked running long-running agents runnig similations on my laptop, trying to break the app. it was fun looking at what wild ideas agent had to crash the app or break UI.
Where was it published?
X / Twitter
Should the case study embed the video?
Yes — use this link (write below)
first: https://x.com/meldiron/status/2103925364766695512 , then second: https://x.com/meldiron/status/2105038700082319801 , then third: https://x.com/meldiron/status/2106351089477890186
Results & retrospective
How was it received? Who uses it?
Numbers are small on GitHub (25 DMG, 79 zip, 10 stars). Feedback from colleagues or the community? The story behind the “intentful” fork?
Me and my friends, few open source community members
What did you learn?
Publishing macOS apps, and building with native Switch, alongside modern MacOS features, including glassy design
What would you do differently?
Video music doesnt have good quality, I should be learning about that sphere a bit more before making new promo videos.
What’s next for Meraline?
Still active? Landing page, App Store version, iOS, a Decision mode trailer…
Dictation, web landing page, and possibly iOS app.
Anything else I got wrong or should know?
Please include link to X threads I made, I think there was 3 or 4. it includes very good threads with screenshot, and first post always have video.
Screenshot notes
Notes I left on individual screenshots for the captions.
01-app-icon-light.png: Favicon