<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Rebell Yell</title>
    <description>The latest articles on DEV Community by Rebell Yell (@indignity).</description>
    <link>https://dev.to/indignity</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3875655%2Fe14a6d48-113b-4d9d-8efb-fa0b7e5b43b6.png</url>
      <title>DEV Community: Rebell Yell</title>
      <link>https://dev.to/indignity</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/indignity"/>
    <language>en</language>
    <item>
      <title>Whisper on Windows with nothing to install: how a desktop app ships whisper.cpp</title>
      <dc:creator>Rebell Yell</dc:creator>
      <pubDate>Tue, 17 Nov 2026 02:54:33 +0000</pubDate>
      <link>https://dev.to/indignity/whisper-on-windows-with-nothing-to-install-how-a-desktop-app-ships-whispercpp-4fo7</link>
      <guid>https://dev.to/indignity/whisper-on-windows-with-nothing-to-install-how-a-desktop-app-ships-whispercpp-4fo7</guid>
      <description>&lt;p&gt;If you've ever tried to run OpenAI's Whisper on a Windows PC, you know the drill: install Python, &lt;code&gt;pip install openai-whisper&lt;/code&gt;, discover it's running on the CPU, uninstall torch, reinstall it from the CUDA index, fight with versions. Fine for a developer. A wall for the people who actually need transcripts: a student with a recorded class, someone with a two-hour meeting, a journalist with an interview.&lt;/p&gt;

&lt;p&gt;I make &lt;a href="https://appgrabbit.com/?utm_source=devto&amp;amp;utm_medium=post&amp;amp;utm_campaign=whisper-windows" rel="noopener noreferrer"&gt;AppGrabbit&lt;/a&gt;, a Windows app that downloads and transcribes videos. When I added local transcription, the rule was: &lt;strong&gt;the user clicks one button and it works&lt;/strong&gt;. No Python, no CUDA toolkit, no terminal, and the audio never leaves the PC.&lt;/p&gt;

&lt;p&gt;Here's how it's built, with the real numbers.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. whisper.cpp instead of Python
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/ggerganov/whisper.cpp" rel="noopener noreferrer"&gt;whisper.cpp&lt;/a&gt; is Georgi Gerganov's C/C++ port of Whisper, and it publishes official Windows builds:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a CPU build: &lt;code&gt;whisper-bin-x64.zip&lt;/code&gt;, &lt;strong&gt;8.6 MB&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;a CUDA 12.4 build: &lt;code&gt;whisper-cublas-12.4.0-bin-x64.zip&lt;/code&gt;, &lt;strong&gt;674 MB&lt;/strong&gt; (the cuBLAS DLLs are most of it; about 1.2 GB unpacked)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;No Python runtime, no PyTorch: one executable (&lt;code&gt;whisper-cli.exe&lt;/code&gt;) plus its DLLs.&lt;/p&gt;

&lt;p&gt;The app pins a single version (&lt;code&gt;b5130&lt;/code&gt;). Everything is tested against it, and a &lt;code&gt;VERSION&lt;/code&gt; file next to the binaries tells the app when an engine on disk doesn't match, so it gets replaced instead of half-working.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Download on demand, verify, swap in atomically
&lt;/h2&gt;

&lt;p&gt;Nothing Whisper-related ships in the installer. It stays small, and people who only download videos never pay for 1+ GB they won't use.&lt;/p&gt;

&lt;p&gt;When someone turns on transcription:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Only what that PC can use is downloaded:&lt;/strong&gt; the CUDA build if there's a usable NVIDIA card, the CPU build otherwise, plus one model and the voice detector.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every file has its SHA-256 pinned in the app.&lt;/strong&gt; They come from a mirror of the official files in our own GitHub release, and anything that doesn't match is thrown away.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disk space is checked first:&lt;/strong&gt; download size + unpacked size + a 200 MB margin.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It unpacks into a temp folder and swaps it in,&lt;/strong&gt; so a crash halfway never leaves a broken engine behind.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Most of the zip is dropped.&lt;/strong&gt; It contains 25 executables (server, stream, talk-llama, benchmarks…). We keep &lt;code&gt;whisper-cli.exe&lt;/code&gt; and only the DLLs it loads. Fewer unknown &lt;code&gt;.exe&lt;/code&gt; files means fewer antivirus false positives.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The models are quantized GGML files:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;In the app&lt;/th&gt;
&lt;th&gt;File&lt;/th&gt;
&lt;th&gt;Size&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Fast&lt;/td&gt;
&lt;td&gt;&lt;code&gt;ggml-small-q5_1.bin&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;190 MB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Accurate&lt;/td&gt;
&lt;td&gt;&lt;code&gt;ggml-large-v3-turbo-q5_0.bin&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;574 MB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Voice detector&lt;/td&gt;
&lt;td&gt;&lt;code&gt;ggml-silero-v5.1.2.bin&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;0.9 MB&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  3. GPU or CPU, without asking the user
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;whisper_backend&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;gpu&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;gpu_info&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;                     &lt;span class="c1"&gt;# nvidia-smi first, WMI as a fallback&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;gpu&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;vendor&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="s2"&gt;"nvidia"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="s2"&gt;"cpu"&lt;/span&gt;                     &lt;span class="c1"&gt;# AMD / Intel run the CPU build&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;gpu&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;compute_cap&lt;/span&gt; &lt;span class="k"&gt;is&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="kc"&gt;None&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;gpu&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;compute_cap&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mf"&gt;5.0&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="s2"&gt;"cpu"&lt;/span&gt;                     &lt;span class="c1"&gt;# Kepler and older can't run CUDA 12&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;gpu&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;driver&lt;/span&gt; &lt;span class="k"&gt;is&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="kc"&gt;None&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;gpu&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;driver&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mf"&gt;551.61&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="s2"&gt;"cpu"&lt;/span&gt;                     &lt;span class="c1"&gt;# CUDA 12.4 needs this driver on Windows&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;settings&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"whisper_cuda_failed"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="s2"&gt;"cpu"&lt;/span&gt;                     &lt;span class="c1"&gt;# the GPU failed its smoke test once&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="s2"&gt;"cuda"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Things that bit us:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;nvidia-smi --query-gpu=name,driver_version,memory.total&lt;/code&gt; gives the real driver version. Compute capability goes in a &lt;strong&gt;separate&lt;/strong&gt; query, because older &lt;code&gt;nvidia-smi&lt;/code&gt; versions don't know &lt;code&gt;compute_cap&lt;/code&gt; and would fail the whole call.&lt;/li&gt;
&lt;li&gt;With no &lt;code&gt;nvidia-smi&lt;/code&gt;, the app falls back to WMI (&lt;code&gt;Win32_VideoController&lt;/code&gt;) and skips remote-desktop and virtual adapters, which are not real cards.&lt;/li&gt;
&lt;li&gt;On some laptops the query takes up to 25 seconds, so the result is cached for an hour and never re-queried while a job is running.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  4. A smoke test before the first real job
&lt;/h2&gt;

&lt;p&gt;After installing, the app runs &lt;code&gt;whisper-cli&lt;/code&gt; on &lt;strong&gt;3 seconds of silence&lt;/strong&gt;: a 16 kHz mono WAV written with Python's &lt;code&gt;wave&lt;/code&gt; module, so not even FFmpeg is involved.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;If the CUDA build exits with an error, the app installs the CPU build and runs the same test. If that one passes, the GPU is to blame: the app remembers it, stays on the CPU, and deletes the 1.2 GB &lt;code&gt;cuda/&lt;/code&gt; folder.&lt;/li&gt;
&lt;li&gt;Only a real non-zero exit from &lt;code&gt;whisper-cli&lt;/code&gt; counts against the GPU. A timeout or a missing file is an environment problem, and the card is not blamed for it.&lt;/li&gt;
&lt;li&gt;The test waits if another transcription is using the GPU, because sharing it would fail on memory and look like a broken card.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  5. The actual call
&lt;/h2&gt;

&lt;p&gt;FFmpeg, already bundled for downloads, converts the audio to 16 kHz mono WAV. Then:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;whisper-cli -m ggml-large-v3-turbo-q5_0.bin -f audio.wav -l auto -t 4 ^
  --vad --vad-model ggml-silero-v5.1.2.bin -osrt -ojf -of output -pp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;-l auto&lt;/code&gt; detects the language. The user can also pin Spanish, English or Portuguese.&lt;/li&gt;
&lt;li&gt;Threads: &lt;code&gt;min(8, cores)&lt;/code&gt; on the CPU, 4 on the GPU.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;--vad&lt;/code&gt; runs the Silero voice detector, which skips music and silence. Without it, large-v3-turbo &lt;strong&gt;invented a credits line&lt;/strong&gt; ("Ana Luiz Revisor: Claudia Sander") over the opening music of a TEDx talk, a known Whisper habit picked up from subtitle training data.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;-osrt&lt;/code&gt; writes subtitles, and &lt;code&gt;-ojf&lt;/code&gt; writes full JSON with per-token confidence.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;-pp&lt;/code&gt; prints progress, which the app parses into a real percentage and ETA.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  6. Honest numbers
&lt;/h2&gt;

&lt;p&gt;Measured on a laptop: Ryzen 7 6800HS (8 cores), RTX 3050 Laptop with 4 GB, Windows 11. The test was a Spanish TEDx talk (7:19), compared against the human transcript. The word error rate (WER) was measured after normalizing case, accents and punctuation.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;CPU (8 threads)&lt;/th&gt;
&lt;th&gt;RTX 3050&lt;/th&gt;
&lt;th&gt;Word error rate&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;small q5_1&lt;/td&gt;
&lt;td&gt;81 s&lt;/td&gt;
&lt;td&gt;21 s&lt;/td&gt;
&lt;td&gt;8.5 %&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;medium q5_0&lt;/td&gt;
&lt;td&gt;208 s&lt;/td&gt;
&lt;td&gt;34 s&lt;/td&gt;
&lt;td&gt;5.9 %&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;large-v3-turbo q5_0&lt;/td&gt;
&lt;td&gt;276 s&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;19 s&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;5.1 %&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;On an English NASA video (5:31), turbo took 13 seconds on the GPU, at 2.6–3.2 % WER.&lt;/p&gt;

&lt;p&gt;What we took from it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;With a GPU, turbo is both the fastest and the most accurate&lt;/strong&gt;, so it's the default whenever a usable NVIDIA card is found.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Without a GPU, small is the only sane default&lt;/strong&gt;: about 2 minutes per 10 minutes of audio. Turbo on the CPU takes about 6 minutes per 10.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Medium is never the best choice.&lt;/strong&gt; It's slower than turbo on the GPU and less accurate everywhere, so the app ships only small and turbo.&lt;/li&gt;
&lt;li&gt;Most of small's Spanish errors are accents and spelling ("¿qué hacés?" becomes "¿qué haces?"), not invented words. The meaning survives.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  7. What it doesn't do
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;No speaker labels.&lt;/li&gt;
&lt;li&gt;No translation.&lt;/li&gt;
&lt;li&gt;No live microphone.&lt;/li&gt;
&lt;li&gt;AMD and Intel GPUs run the CPU build for now.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  From the user's side
&lt;/h2&gt;

&lt;p&gt;All of the above collapses into one download button in Settings and this screen: paste a link or drop your own file, pick a prompt for your AI if you want one, and you get a clean &lt;code&gt;.txt&lt;/code&gt; plus a timed &lt;code&gt;.srt&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh0yl3nv9j6gw9svjtqa7.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh0yl3nv9j6gw9svjtqa7.png" alt="AppGrabbit's transcript tab on Windows: your own file, language, and a prompt for your AI" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If you want to see it in action, here's &lt;a href="https://appgrabbit.com/whisper-windows-no-install/?utm_source=devto&amp;amp;utm_medium=post&amp;amp;utm_campaign=whisper-windows" rel="noopener noreferrer"&gt;Whisper on Windows with nothing to install&lt;/a&gt;. AppGrabbit is free to start (5 credits a month) on Windows 10 and 11.&lt;/p&gt;

&lt;p&gt;Questions about the GPU detection, the VAD setup or the numbers? Happy to answer in the comments.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;— CrowSo Studio&lt;/em&gt;&lt;/p&gt;

</description>
      <category>whisper</category>
      <category>windows</category>
      <category>ai</category>
      <category>showdev</category>
    </item>
    <item>
      <title>I rebuilt my Windows video downloader from scratch — yt-dlp + FFmpeg + PyInstaller, 1,700 downloads later (now AppGrabbit)</title>
      <dc:creator>Rebell Yell</dc:creator>
      <pubDate>Fri, 10 Jul 2026 19:01:14 +0000</pubDate>
      <link>https://dev.to/indignity/i-rebuilt-my-windows-video-downloader-from-scratch-yt-dlp-ffmpeg-pyinstaller-1700-downloads-1f0l</link>
      <guid>https://dev.to/indignity/i-rebuilt-my-windows-video-downloader-from-scratch-yt-dlp-ffmpeg-pyinstaller-1700-downloads-1f0l</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Update, September 2026.&lt;/strong&gt; Grabbit is now &lt;strong&gt;AppGrabbit&lt;/strong&gt; (version 3.2.0) and a few things below are out of date. The Free plan moved from per-feature caps to &lt;strong&gt;5 credits a month&lt;/strong&gt; (a video, a batch of up to 10 links, a transcript or a converted file is 1 credit; claim 1 more every 48 hours), so the converter is no longer Pro-only. Transcripts no longer depend on YouTube captions: videos without them, and your own files (a recorded class, a meeting, a voice note), are &lt;strong&gt;transcribed locally with whisper.cpp&lt;/strong&gt;, on an NVIDIA card or the CPU, nothing uploaded. Pricing is &lt;strong&gt;USD 8.99 a month, monthly only (the annual plan was dropped)&lt;/strong&gt;, or a &lt;strong&gt;one-time 7-day pass for USD 3.99&lt;/strong&gt;. Details: &lt;a href="https://appgrabbit.com/blog/whats-new-in-appgrabbit-3-2/" rel="noopener noreferrer"&gt;what's new in 3.2&lt;/a&gt; · &lt;a href="https://appgrabbit.com/whisper-windows-no-install/" rel="noopener noreferrer"&gt;how the local Whisper setup works&lt;/a&gt; · &lt;a href="https://appgrabbit.com/blog/how-credits-work/" rel="noopener noreferrer"&gt;how credits work&lt;/a&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Three months ago I posted here about Grabbit, a desktop video downloader I built for Windows. The app has grown a lot since then — new engine, local file converter, batch mode, transcript extraction — so I wanted to write a proper update with the actual technical details.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Grabbit does (2026 edition)
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://appgrabbit.com" rel="noopener noreferrer"&gt;Grabbit&lt;/a&gt;&lt;/strong&gt; is a Windows 10/11 desktop app that downloads videos from 8+ platforms — YouTube, TikTok, Instagram, Facebook, Twitter/X, Pinterest, Twitch, SoundCloud — up to 4K quality. Everything runs locally. No cloud, no account required, no file upload anywhere.&lt;/p&gt;

&lt;p&gt;It also ships with a &lt;strong&gt;local FFmpeg converter&lt;/strong&gt; that converts between video and image formats entirely on your machine (Pro-only when this was written; open to Free since 3.2.0, 1 credit per file).&lt;/p&gt;

&lt;p&gt;Supported downloads: &lt;a href="https://appgrabbit.com/youtube-to-mp3/" rel="noopener noreferrer"&gt;YouTube to MP3&lt;/a&gt;, &lt;a href="https://appgrabbit.com/tiktok-video-downloader/" rel="noopener noreferrer"&gt;TikTok&lt;/a&gt;, &lt;a href="https://appgrabbit.com/instagram-video-downloader/" rel="noopener noreferrer"&gt;Instagram Reels&lt;/a&gt;, &lt;a href="https://appgrabbit.com/pinterest-video-downloader/" rel="noopener noreferrer"&gt;Pinterest&lt;/a&gt;, &lt;a href="https://appgrabbit.com/youtube-shorts-downloader/" rel="noopener noreferrer"&gt;YouTube Shorts&lt;/a&gt;, &lt;a href="https://appgrabbit.com/soundcloud-downloader/" rel="noopener noreferrer"&gt;SoundCloud&lt;/a&gt;, and more.&lt;/p&gt;

&lt;p&gt;Supported conversions: &lt;a href="https://appgrabbit.com/video-to-mp3/" rel="noopener noreferrer"&gt;Video → MP3&lt;/a&gt;, &lt;a href="https://appgrabbit.com/webm-to-mp4/" rel="noopener noreferrer"&gt;WebM → MP4&lt;/a&gt;, &lt;a href="https://appgrabbit.com/heic-to-jpg/" rel="noopener noreferrer"&gt;HEIC → JPG&lt;/a&gt;, &lt;a href="https://appgrabbit.com/webp-to-jpg/" rel="noopener noreferrer"&gt;WebP → JPG&lt;/a&gt;, &lt;a href="https://appgrabbit.com/video-to-prores/" rel="noopener noreferrer"&gt;Video → ProRes HQ&lt;/a&gt;, &lt;a href="https://appgrabbit.com/video-to-dnxhr/" rel="noopener noreferrer"&gt;Video → DNxHR HQ&lt;/a&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  The stack (unchanged from v1, for a reason)
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Backend:      Python + Flask + yt-dlp + FFmpeg
Frontend:     Vanilla JS (no framework)
Auth/DB:      Supabase (PostgreSQL + Edge Functions)
Payments:     Stripe → Supabase Edge Function → Resend
Installer:    PyInstaller + Inno Setup
Distribution: GitHub Releases (auto-update via version check)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I kept the same stack because it works. The temptation to rewrite in Electron or Tauri was real but I resisted it. Adding complexity wouldn't have shipped any of the features below.&lt;/p&gt;




&lt;h2&gt;
  
  
  What's new: the FFmpeg Converter
&lt;/h2&gt;

&lt;p&gt;The biggest addition since April is the local Converter. Users can drop a local file (MP4, MKV, AVI, MOV, WebM, HEIC, WebP) and convert it to another format — entirely on their machine. No upload, no size limit.&lt;/p&gt;

&lt;p&gt;The trickiest part: &lt;strong&gt;bundling FFmpeg and pillow-heif with PyInstaller&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;For FFmpeg I use &lt;code&gt;imageio-ffmpeg&lt;/code&gt; which ships its own binary:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;imageio_ffmpeg&lt;/span&gt;
&lt;span class="n"&gt;ffmpeg_path&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;imageio_ffmpeg&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get_ffmpeg_exe&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This gives you a reliable path whether you're running from source or from a frozen PyInstaller bundle. No &lt;code&gt;which ffmpeg&lt;/code&gt; nonsense, no PATH issues on Windows.&lt;/p&gt;

&lt;p&gt;For HEIC support (&lt;code&gt;pillow-heif&lt;/code&gt;), PyInstaller doesn't collect the native DLLs automatically. You need to force it in the &lt;code&gt;.spec&lt;/code&gt; file:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# grabbit.spec
&lt;/span&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;PyInstaller.utils.hooks&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;collect_all&lt;/span&gt;
&lt;span class="n"&gt;datas_heif&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;binaries_heif&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;hiddenimports_heif&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;collect_all&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;pillow_heif&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Without that, HEIC conversion silently fails at runtime on the packaged exe — no error, just nothing happens. Took an embarrassing amount of time to debug.&lt;/p&gt;




&lt;h2&gt;
  
  
  What's new: Batch mode
&lt;/h2&gt;

&lt;p&gt;Users can paste a list of URLs — one per line — and Grabbit queues them all (Free batches are capped at 10 links per batch; Pro has no cap). The implementation is straightforward but the UX took iteration: users need to know which URLs failed and why, without it feeling like a wall of errors.&lt;/p&gt;

&lt;p&gt;I ended up with a per-URL status row that shows &lt;code&gt;queued → downloading → done/error&lt;/code&gt; in real time. The progress streaming still holds — polling every 600ms against a lightweight endpoint works well enough that I haven't needed WebSockets.&lt;/p&gt;




&lt;h2&gt;
  
  
  What's new: Transcript mode
&lt;/h2&gt;

&lt;p&gt;Grabbit can extract transcripts from YouTube videos — useful for summarizing, repurposing content, or accessibility. It uses yt-dlp's subtitle extraction under the hood:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;ydl_opts&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;writesubtitles&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;writeautomaticsub&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;subtitleslangs&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;en&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
    &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;skip_download&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Auto-generated captions aren't perfect but for most YouTube content they're good enough for summarization use cases. &lt;em&gt;(Since 3.2.0 this is only the fast path: when a video has no captions, or when you drop a file of your own, the audio is transcribed locally with whisper.cpp; see the update at the top.)&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  What's new: self-updating yt-dlp
&lt;/h2&gt;

&lt;p&gt;yt-dlp breaks constantly as platforms change their APIs. Rather than shipping a new app version every time, Grabbit updates yt-dlp silently in the background on launch:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;subprocess&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;sys&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;update_ytdlp&lt;/span&gt;&lt;span class="p"&gt;():&lt;/span&gt;
    &lt;span class="n"&gt;subprocess&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;sys&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;executable&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;-m&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;pip&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;install&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;--upgrade&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;yt-dlp&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
        &lt;span class="n"&gt;capture_output&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When frozen with PyInstaller, &lt;code&gt;sys.executable&lt;/code&gt; points to the bundled Python — so pip updates the bundled yt-dlp, not a system one. This means users always have a working downloader even when a platform changes something overnight.&lt;/p&gt;




&lt;h2&gt;
  
  
  Numbers after 3 months
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;1,700+ installs&lt;/strong&gt; via GitHub Releases&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Top countries:&lt;/strong&gt; Mexico, US, Peru, Spain, Argentina, Colombia&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Top traffic sources:&lt;/strong&gt; Direct, Google Organic, ChatGPT referral&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free plan (July 2026):&lt;/strong&gt; 10 single downloads, 1 batch (up to 10 links), 2 transcripts — replaced in 3.2.0 by 5 credits a month&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pro:&lt;/strong&gt; $8.99/month, monthly only (the annual plan was dropped) — unlimited downloads and batches, playlists, all converter formats&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The LATAM tilt was unexpected but makes sense in hindsight — there's a real gap for a clean Spanish-language download tool. I shipped 19 new SEO landing pages in Spanish last week targeting that.&lt;/p&gt;




&lt;h2&gt;
  
  
  What I'd do differently
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Ship the Chrome Extension earlier.&lt;/strong&gt; The extension puts a one-click Grab button directly on video pages while you're watching. It's one of the highest-retention features and I waited too long to build it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Don't fight PyInstaller — embrace the spec file.&lt;/strong&gt; I spent weeks trying to make things work without touching the &lt;code&gt;.spec&lt;/code&gt;. The right answer is always: open the spec, add what you need, move on.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Ignore the "just use Electron" advice.&lt;/strong&gt; Flask + Vanilla JS served from localhost is genuinely fine for a tool like this. The bundle is smaller, startup is faster, and I understand every line of it.&lt;/p&gt;




&lt;p&gt;If you're building something similar — local desktop tool, Python backend, freemium model — happy to answer questions in the comments.&lt;/p&gt;

&lt;p&gt;If you want to try it: &lt;a href="https://appgrabbit.com" rel="noopener noreferrer"&gt;appgrabbit.com&lt;/a&gt; — Windows only, free plan available, no account required.&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>python</category>
      <category>ffmpeg</category>
      <category>windows</category>
    </item>
    <item>
      <title>I built a free desktop video downloader for Windows — Grabbit (now AppGrabbit)</title>
      <dc:creator>Rebell Yell</dc:creator>
      <pubDate>Mon, 13 Apr 2026 00:40:24 +0000</pubDate>
      <link>https://dev.to/indignity/i-built-a-free-desktop-video-downloader-for-windows-grabbit-3knc</link>
      <guid>https://dev.to/indignity/i-built-a-free-desktop-video-downloader-for-windows-grabbit-3knc</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Update, September 2026.&lt;/strong&gt; Grabbit is now &lt;strong&gt;AppGrabbit&lt;/strong&gt; (version 3.2.0). The Free plan no longer has a daily cap: it runs on &lt;strong&gt;5 credits a month&lt;/strong&gt; (a video, a batch, a transcript or a converted file is 1 credit) and you can claim 1 more every 48 hours. The app now &lt;strong&gt;transcribes any video, with or without captions, on your PC&lt;/strong&gt; with whisper.cpp, and your own recordings too. Pricing is &lt;strong&gt;USD 8.99 a month, monthly only (no annual plan)&lt;/strong&gt;, or a &lt;strong&gt;one-time 7-day pass for USD 3.99&lt;/strong&gt;. Everything that changed: &lt;a href="https://appgrabbit.com/blog/whats-new-in-appgrabbit-3-2/" rel="noopener noreferrer"&gt;what's new in 3.2&lt;/a&gt;. The rest of this post is the original April story, kept as it was.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I've been wanting to download videos from YouTube and TikTok for a long time, but every solution I found was either a sketchy website full of ads, a command-line tool that non-technical users couldn't use, or a bloated app that required an account and uploaded your files to a cloud server.&lt;br&gt;
So I built my own.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What it does&lt;/strong&gt;&lt;br&gt;
Grabbit is a desktop app for Windows that downloads videos from YouTube, TikTok, Instagram, Facebook, Twitter, Pinterest, Twitch and SoundCloud — up to 4K/8K quality. It runs a local Flask server and opens in your browser, so the UI is just a web page served from localhost:5000. No Electron, no heavy framework. Just Python and Vanilla JS.&lt;br&gt;
It also has a Chrome Extension that puts a Grab button directly on video pages. While you're watching a video, click Grab — it gets added to a download queue in real time. No copying links, no switching tabs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How I built it&lt;/strong&gt;&lt;br&gt;
The stack is simple on purpose:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Backend:&lt;/strong&gt; Python + Flask + yt-dlp + ffmpeg&lt;br&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; Vanilla JS SPA (no React, no Vue)&lt;br&gt;
&lt;strong&gt;License system:&lt;/strong&gt; Supabase (PostgreSQL + Edge Functions)&lt;br&gt;
&lt;strong&gt;Payments:&lt;/strong&gt; Stripe webhooks → Supabase Edge Function → Resend email&lt;br&gt;
&lt;strong&gt;Installer:&lt;/strong&gt; PyInstaller + Inno Setup&lt;br&gt;
&lt;strong&gt;Distribution:&lt;/strong&gt; GitHub Releases&lt;/p&gt;

&lt;p&gt;The trickiest part was getting the download progress to stream in real time. yt-dlp outputs progress line by line, so I parse each line with regex and push it to a polling endpoint the frontend checks every 600ms. Not the most elegant solution but it works reliably.&lt;br&gt;
Another challenge was packaging with PyInstaller. When frozen, &lt;strong&gt;file&lt;/strong&gt; points inside the bundle temp directory, not where the exe lives. So I had to detect sys.frozen and use sys.executable to find the actual working directory. Took me longer than I'd like to admit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I learned&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Building a freemium desktop app from scratch taught me a lot about things I'd never dealt with before — code signing certificates, Windows Smart App Control, subprocess encoding issues on Windows (cp1252 vs utf-8), auto-update architecture without breaking user data, and how to structure a license system that works offline when Supabase is unreachable.&lt;br&gt;
The most important thing I learned: ship early. The first version was rough. But shipping it gave me real feedback and the motivation to keep going.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where it is now&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Grabbit is live. Free plan included. Pro plan at $8.99/month (monthly only, no annual plan) for unlimited downloads, 4K/8K, all platforms and segment download. &lt;em&gt;(Since September 2026 the Free plan runs on credits and the app is called AppGrabbit; see the update at the top.)&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;If you're curious: &lt;a href="https://appgrabbit.com" rel="noopener noreferrer"&gt;appgrabbit.com&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Happy to answer any questions about the stack or implementation details.&lt;br&gt;
If you try it, let me know what you think — feedback is welcome.&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>javascript</category>
      <category>python</category>
      <category>showdev</category>
    </item>
  </channel>
</rss>
