The take you already nailed
The voice memo where you explained it perfectly IS the content. This skill skips the script-and-re-record loop entirely — the recording becomes the video.
Your voice in, a captioned short out
A voice memo, a podcast segment, a riff recorded on your phone — Claude stages it, ViralMint transcribes it locally with Whisper, and the video is built around your audio: word-timed captions, b-roll matched to what you're saying, optional music. No TTS, no re-recording.
You type
Turn this voice memo into a vertical short: ~/Desktop/riff.m4a
…and Claude drives the whole pipeline below on your machine.
Why this beats a blank script box
No TTS bill, no TTS feel
Audio mode charges no text-to-speech because there is none — real human delivery, which the algorithm and the audience both prefer.
Captions that actually sync
Because the same local Whisper pass that transcribes the clip also times the captions, every word lands on screen as it's spoken — not drifting subtitle blocks.
How Claude turns audio into a captioned short
-
Stage the audio
stage_audio_for_smart_video normalizes the clip and transcribes it once with local Whisper — returning the transcript so Claude can sanity-check it before anything renders.
-
Your voice is the voice track
generate_smart_video takes the staged audio instead of a script. No TTS is billed — the clip you recorded is what plays.
-
Captions and b-roll follow the words
Whisper's word timestamps drive word-by-word animated captions, and the transcript drives scene retrieval, so the footage matches what's being said.
-
Collect the short
The finished 9:16 mp4 lands in your library with AI-drafted title, description and tags — ready to post.
MCP tools it calls:
stage_audio_for_smart_videogenerate_smart_videowait_for_joblist_videos Connect it in 4 steps
-
Install ViralMint
Download the free, open-source desktop app (macOS / Windows / Linux) and open it. It bundles the whole pipeline — yt-dlp, Whisper, FFmpeg — and the MCP server.
-
Open the in-app MCP page
It shows your loopback endpoint and a bearer token, with a ready-to-paste Claude Code command and a JSON config snippet for Cursor / Claude Desktop.
-
Add the server to your agent
Paste the command. Your client loads ViralMint's tools, the built-in workflow recipes (this skill is one) and read-only resources like your balance and recent videos.
-
Ask for the outcome
Type the prompt above. Claude runs the recipe end-to-end against the local app and hands you the finished mp4 — no clicking through the UI.
ViralMint skills vs typical agent video tools
| What matters | Typical paid agent tools | ViralMint |
|---|---|---|
| Agent tools / skills | Behind a paid tier | Free & open-source (AGPL) |
| Where it runs | Their cloud | Your machine (loopback MCP) |
| Your footage & data | Uploaded to their servers | Stays local — only AI calls leave |
| Pricing | Monthly subscription | Prepaid, pay-per-render at ~cost |
| Output watermark | On free tiers | Never |
| Script quality | Prompt → generic script | Grounded in real breakouts + your data |
highlighted column = clearer fit.
FAQ
Frequently asked
What audio formats and lengths work?
Common audio formats (m4a, mp3, wav) and even video files — the audio track is extracted. Caps: 10 minutes and 200 MB per staged clip. For a full podcast episode, use the Long Video → Clips skill instead to mine multiple shorts.
Does my recording get uploaded for transcription?
No — transcription runs locally with Whisper on your machine. The audio is normalized and staged on-device; only the b-roll retrieval and optional AI extras touch the cloud.
What if the recording is noisy?
Claude checks the transcript after staging — if it's garbled, it runs the audio enhancer (denoise + loudness normalization) first and re-stages, so the captions are built from a clean read.
Can I use my own footage instead of stock b-roll?
Yes — import your clips for assembly and they're blended with (or replace) stock. AI images and AI video clips can also be mixed in for a premium look, billed per generation with a quote first.
Make {skill.name.toLowerCase()} from a chat
Install the free ViralMint desktop app, connect it to Claude Code or Cursor, and run this skill — no subscription, no watermark, your footage never leaves your machine.