Free AI App Builder: Talk to It, No Signup

The feature none of the other free builders talk about

Search results for "free ai app builder" all converge on the same story: pick a template, sign up, drag a few components, and the free tier ends when you want to publish. The tools are text first and signup first. They rarely mention voice.

mk0r puts a microphone button right next to the prompt box. Press it, say what you want to build, and the transcript drops into the input field. You can edit it, add to it, or submit it as-is. This is not a paid add-on. It ships on the free anonymous path along with everything else.

What happens when you hold the mic

The voice path is a small stack of decisions, each one aimed at getting a clean transcript to the model fast. You can read the whole thing in src/hooks/useVoiceInput.ts and src/app/api/transcribe/route.ts. Here is the path a single recording takes:

1. Capture. navigator.mediaDevices.getUserMedia requests the mic with echoCancellation, noiseSuppression, and autoGainControl turned on.
2. Encode. MediaRecorder negotiates the best mime type available, in order: audio/webm;codecs=opus, then webm, then mp4 as a Safari fallback. It records in 100 ms chunks so nothing is buffered longer than it needs to be.
3. Ship. On stop, chunks are stitched into a single Blob and POSTed to /api/transcribe with the mime type preserved in the content-type header.
4. Transcribe. The server forwards the raw audio body straight to Deepgram's nova-2 model with smart_format=true and punctuate=true. No buffering, no disk writes, just a streamed proxy.
5. Review. The transcript returns as plain text and lands in the prompt input. You get to see it before you submit. Voice is not a bypass around the prompt; it fills the prompt.

That last point matters. Most voice-first tools send your words to the model immediately and you discover misheard words in the output. mk0r keeps the review step because transcripts are wrong sometimes, and prompts compound those errors into bad apps.

Why nova-2 and not the browser's built-in recognition

The Web Speech API exists and is free, so why route through a server at all? Two reasons. The first is quality: nova-2 handles technical vocabulary, code-adjacent terms, and accent variance better than the default browser recognizer on most platforms. When you are describing an app you end up saying things like "radix primitives" or "Postgres schema" and consumer speech recognition tends to mangle those.

The second is consistency. Web Speech is implemented differently in Chrome, Safari, and Firefox, and in some browsers it phones home to cloud services you have no visibility into. Routing to a single provider keeps behavior predictable across devices and keeps the formatting (capitalization, punctuation) uniform so the prompt is usable as-is.

The rest of the free surface

Voice input is the headline, but it is one of several things mk0r hands you without asking for anything in return:

• No signup. Session is a client-generated UUID in sessionStorage.
• Two generation modes. Quick mode streams a single HTML file from Claude Haiku. VM mode spins a Freestyle sandbox with Vite, React, Chromium, and Playwright.
• Browser-verified output. In VM mode the agent opens the app in a real Chromium instance and checks the rendered page before showing you the preview.
• Snapshot-cached boots. VMs start from a persisted snapshot in 2 to 3 seconds after the first build.
• Iterative follow-ups. Each message in the same session edits the same app; you are not starting fresh each prompt.

All of these are on the anonymous free path. There is no tier above them gating basic usage.

A practical example

A builder I was watching the other day opened mk0r, held the mic, and said: "a kanban board with three columns where I can drag cards between them and edit card text inline, save to local storage."

The transcript came back cleanly with the commas in the right places. They pressed Enter. The VM booted, the agent wrote a React app with dnd-kit and localStorage persistence, Chromium rendered it, the agent dragged a card to confirm the behavior worked, and the preview appeared. No signup screen, no pricing page, no input mode toggle.

This is the workflow the page is about. Not the idea of a free app builder, the actual sequence of a person talking to a tool and getting a thing they can use.