# 1 Year of Building a "Bring Your Own API Key" Transcription App — and Why It's Now an Automation Hub > Source: > Published: 2026-10-09 20:48:10+00:00 Hi, I'm a solo iOS developer. About a year ago, when I first released my transcription app, I wrote this: "The moment you go subscription, you can't stop." "Even if only one person has ever paid, you have to keep the servers running — even at a loss." "A single traffic spike can degrade your service." While every AI app was racing toward monthly subscriptions, I realized that for a solo developer, running a backend long-term is just too much risk. Then it hit me: **What if I just build the "vessel" — and let the user bring their own OpenAI API key?** A year later, WhisperDirect has grown into something with transcription, summarization, meeting minutes, and external automation built in. This post is the story of that year. Why an "API-direct, fully one-time-purchase" transcription app ended up here. WhisperDirect started with a very simple idea: In other words, WhisperDirect was, from day one, **a vessel for using OpenAI via your own API key**. No backend to maintain, no middleman markup. You pay OpenAI for what you use. That part of the design has never changed. On the other hand, I had another app: **WhisText**. Think of it like Fedora vs. Red Hat in the Linux world. Cutting-edge tech goes into **WhisText** first. Whatever survives real-world use gets ported into **WhisperDirect**. Over the past year, that cycle has accelerated dramatically. WhisText's transcription originally ran on my own GPU server. The model was Whisper large-v3-turbo. My plan was: let people use it for free, watch the usage, then figure out pricing. Reality was less generous. Usage never justified keeping a GPU running 24/7, and the bills kept coming. So I switched to CPU and started testing every ASR engine I could find. My dev server's `asr.sherpa` directory still holds the scars: I tried different language combinations, split audio into chunks for parallel processing, tuned for speed. A lot of unglamorous trial and error. Eventually, **Apple Speech became the center of WhisText's final version.** I had actually tried Apple's native Apple Speech pretty early on. At the time, it was unusable. The biggest blocker was the **1-minute limit**. For an app that records meetings and produces minutes, cutting off every 60 seconds is fatal. So I went back to tuning my own server. Then iOS 26 changed everything. Apple Speech's accuracy jumped, and the old limits and instability were gone. In my own benchmarks, it was **matching or beating** engines I'd been running on my own server — Parakeet, ReazonSpeech, all of them. So WhisText's final version put Apple Speech at the core. No server round-trip means: **the instant you speak into the mic, the text appears.** I added live preview as a bonus. That said, live preview is a bonus. If you want accurate meeting minutes, batch-processing the whole recording with Whisper API afterward still produces cleaner results. Live preview is more about the *feeling* — "it's recording," "AI is working right now." But if Apple Speech on iOS 26 is this good, there must be more we can do on-device. So the modules that moved to on-device in WhisText got ported into WhisperDirect: **These became the foundation that works without an API key, for free.** The accuracy improvements in iOS 26's Apple Speech are what made this foundation actually usable. The latest version — currently in App Store review — packs in a lot more. WhisperDirect is a voice transcription and summarization app built around Whisper API's accuracy. Alongside Whisper API (with your own key), it ships with on-device features: Apple Speech, speaker diarization, and OCR. For summarization and minutes, you can choose your LLM from OpenAI, Gemini, or any OpenAI-compatible endpoint. **The app is a one-time purchase. No subscriptions.** Cost reference: Whisper API is about **$0.006/min**. That's roughly **$0.36/hour** — about **8.7 hours of transcription for around $3**. **With a local LLM, you can bring the summarization / minutes cost down to zero.** For a few thousand characters, cloud APIs cost just a few cents. That kind of freedom is only possible for a solo developer. I built this for myself, honestly. You can POST transcription / summary / minutes results to an external endpoint, each at the moment it completes. From there, **n8n handles the rest.** With n8n, you can do whatever LLM processing you want, then route to Notion, Google Sheets, Slack, or anything else — just configure the URL. Header auth is supported too. The payload looks like this: ``` { "text": "...", "type": "transcript", "recorded_at": "2026-10-09T10:00:00Z", "sent_at": "2026-10-09T10:05:00Z", "source": "whisperdirect", "version": "1.0" } ``` For a company-run app, you need a monthly subscription just to cover server costs, salaries, and ad spend. But WhisperDirect has no backend to maintain. The app is a one-time purchase. When an API is used, the user pays the provider directly — at cost. I can't spend money on ads. So instead of competing on ad spend, I compete on cost-performance and freedom. What I can do as a solo developer, and what I can't. I've been carrying both, the whole way here. The latest version is currently in App Store review (live preview, etc.). There's a 7-day free trial. If this sounds interesting, search for WhisperDirect on the App Store. Once it's approved, give the new transcription experience a try. My recommended setup: The point is: **you don't have to choose between "cheap" and "accurate." You decide per recording, per segment.** WhisperDirect: [https://apps.apple.com/app/id6748595475](https://apps.apple.com/app/id6748595475)