Marcus was 5000 words into a design doc at 11pm, with two more sections to go, when his dictation tool hit the word cap. Back to typing. Fragmented prose. Lost thinking flow. By morning, it wasn't worth committing.
The voice promise wasn't designed for model-era development
Old voice tools pitched speed: talk faster than you type. That was true for code itself. You'd dictate a quick loop, a function signature, variable names. Voice made sense for discrete chunks of syntax.
But engineers don't dictate code anymore. They dictate intent.
Design docs explaining why the system should work a certain way. PR descriptions and review comments walking through architectural decisions. Slack threads explaining bug investigations: what you observed, what you checked, what you ruled out, what you think is happening, what you're testing next. Incident postmortems. Linear issue specifications. The shape of the work became explain-the-spec, not write-the-code. Same tool. Different job. Different volume. Different pace.
The old metric, words per minute, stopped mattering. What mattered was completeness. The ability to express a full thought from start to finish without being interrupted by an arbitrary cap.
The voice tool's selling point, velocity, became irrelevant. What you actually needed was endurance: the ability to dictate the entire thought without hitting a wall mid-sentence.
When the word limit hits mid-flow
This is where the metering kicks in. Wispr charges $14/month and caps the free tier at 5000 words per month. Willow is $12/month with a similar ceiling. Superwhisper is $8.49 with an even lower cap. The free tiers aren't designed around generosity. They're designed around control: use a bit, and you upgrade. Use more, and you upgrade faster.
For Marcus, for any engineer whose design doc runs long, the cap doesn't arrive at a natural stopping point. It arrives mid-thought. Mid-spec. Mid-flow. The friction isn't having to type a little; it's the context switch that kills your momentum the instant you've articulated the most important part of the idea.
You can always type. The voice tool was supposed to free you from typing. Then it forces you back.
Why cloud metering isn't about cost
Here's the part that shouldn't surprise any engineer: the variable cost of transcribing a word to text on the cloud is near-zero. Hosting costs are fixed. Bandwidth is cheap. The constraint is business, not technical.
Cloud-based voice tools meter free tiers not because transcription expenses them. They meter because usage limits are growth levers. You hit the cap. You upgrade. The metering is artificial.
And for developers, there's an additional cost that the pricing doesn't mention: data leaving the device. Your voice drafts travel to servers you don't control. Processed by systems you can't inspect. Code snippets in your design docs. Debug output in your incident narratives. The transcription happens offshore. You have no visibility into where the recordings go, how long they're kept, who sees them, whether they're used to train future models.
For engineers who refuse cloud-based transcription on principle, who won't let code IP leave the device, these tools aren't an option no matter the price.
Local Whisper changes the equation
Recitey runs Whisper locally on your Windows machine. No cloud calls. No variable cost per word. The speech-to-text model sits on your hardware. The processing stays on device. Your words never leave. The IP doesn't leave.
Because there's no variable cost, there's no metering business model to support. So there's no word limit on the free tier. You dictate until the design doc is complete. Until the incident postmortem is finished. Until you've explained the bug investigation fully. Until the thought is done.
The bottleneck, if there is one, is your machine's resources, not someone else's subscription tier. And hardware usually isn't the constraint.
Works in every IDE and app. No lock-in.
Cursor has native dictation. But it's Cursor-only. GitHub has Copilot. But GitHub-only.
Recitey works in Cursor, yes. Also in VS Code via the Claude Code extension. Also in GitHub PR descriptions, Slack threads, Linear tickets, Notion docs, email, system-wide via the clipboard. One speech-to-text layer. Every app you already use.
That matters to developers like Marcus who code in Cursor specifically because tab-complete reduces voice rewrites, but need to write design docs in Notion, file bugs in Sentry, explain incidents in Slack, and post code reviews across GitHub. The tool fits the workflow, not the other way around.
The honest trade-off
This is not for cloud-speed. Cloud voice tools have a latency trick: they stream transcription in real-time while you're still speaking. Words appear in Slack before you've finished the sentence. Recitey doesn't do that. Local processing takes a beat.
This is also not for cloud-rewrite. Some tools let you speak roughly, fragments and fillers and half-thoughts, and the cloud polishes it into a clean sentence. Recitey doesn't do that either. You speak; the tool transcribes.
But if you're Marcus, if you're explaining intent at 11pm without hitting a cap, without leaving code IP to the cloud, without paying $14/month for metering you didn't ask for, then local Whisper becomes invisible. You dictate until you're done. The cap stops being a thing you think about. The thinking becomes the thing you think about.
The bottleneck in model-era development isn't transcription speed. It's explanation depth. Recitey lets you explain without hitting a wall.