Everyone from Instagram influencers to big-name CEOs seems to be saying the same thing: AI is going to be the end of SaaS (Software as a Service).And while I’d love a world with fewer subscriptions, when I look around, I see quite the opposite.If anything, more software subscriptions are popping up every year, often at absurdly inflated prices.
So, I decided to take matters into my own hands and see if I could use AI to replace some of my subscriptions.I took my $20 Claude subscription and tried to vibe code alternatives to many of the tools I use daily.As it turns out, I was able to save a little over $400 a year by running this experiment.
Here are all the subscriptions I replaced using Claude.I saved $144/year on Grammarly Using some clever prompting and Claude Artifacts Close One of the big debates surrounding AI is its use for writing content.I personally find AI-written content very repulsive.
AI-written emails, tweets, and posts often feel unnatural, performative, and manufactured.That said, I do think LLMs have a solid grasp of grammar.While I’ve seen ChatGPT and Claude write meaningless sentences, I’ve rarely (read: never) seen them make grammatical errors.
As such, Grammarly was actually the first subscription I canceled after I started using LLMs.They can effortlessly catch most grammar issues, including spelling mistakes, tense errors, and punctuation problems.Furthermore, since LLMs can sort of “understand” what I’m writing, their suggested corrections are also context-aware.
Say I meant to write “can’t” but accidentally typed “can” instead.Since the sentence is still grammatically valid, Grammarly might let it through.However, an LLM can read the whole passage, realize the sentence contradicts what I’m saying, and flag it.
That said, the typical problem with using an LLM as a Grammarly replacement is the interface.Pasting a passage into a chatbot and getting a corrected version as output is a bad workflow.Grammarly is useful because you can see each error and decide whether to accept each edit.
Fortunately, you can set up a similar workflow using Claude Artifacts.You paste the entire passage you want to proofread into the LLM and ask it to create an Artifact that reproduces the passage, strikes through grammar issues, and displays corrections alongside them.The same approach works with ChatGPT Canvas and Gemini Canvas as well.
Related I turned ChatGPT into a free Grammarly Pro replacement without any vibe coding Your ChatGPT has been Grammarly Pro this whole time—you just didn't ask right.Posts 1 By Dibakar Ghosh I saved another $144/year on Wispr Flow By combining a Python library with an open-source dictation model Close Even though I write for a living, I honestly don’t like typing.Part of it comes down to my posture: I’ve developed wrist pain that makes long writing sessions genuinely painful.
Plus, I never learned to touch-type, and I top out at around 60 WPM (words per minute) when typing, while I speak at roughly 200 WPM.So dictation has always made more sense to me.I used dictation apps long before the AI boom, back when Dragon NaturallySpeaking was the main option.
These days, you’re spoiled for choice, with open-source models like OpenAI’s Whisper and NVIDIA’s Parakeet, plus a bunch of startups packaging them in nicer interfaces and selling access through monthly subscriptions.At the time of writing, Wispr Flow is arguably one of the best speech-to-text tools out there.I used it for a while, and it’s pretty good.
But it’s still a subscription, which means money leaves my pocket every single month.What made it even more frustrating was that it used an open-source model under the hood.So I figured I could build my own tool around these models and give my wallet a bit of a break.
I used Claude to come up with a plan.It recommended RealtimeSTT, an open-source project for real-time transcription, paired with faster-whisper, a faster implementation of the Whisper model.Then I used Claude to vibe code a PowerShell script that combines these tools and launches automatically at system startup.
The script supports both push-to-talk and live transcription.I can press F9, and a small pop-up appears to let me know it's recording.I can then start speaking.
When I press F9 again, it stops recording, transcribes everything I said, and saves both the audio and transcript to a folder.It also copies the transcript to my clipboard and pastes it into whichever text field I have selected.In live transcription mode, I can press F10 and start talking, and the text appears in the active text field as I speak.
You need a reasonably powerful GPU to run transcription models locally at usable speeds.For reference, I’m using an RTX 3060 with 12GB of VRAM, but in my testing, I found 6GB to be good enough.That said, if you’re running a CPU-only setup, NVIDIA Parakeet is supposedly a good option to try out — albeit real-time transcription is going to feel very slow.
I saved another $120 annually on Rize By vibe coding an HTML dashboard and pairing it with ActivityWatch Close I came across Rize fairly recently while taking an online productivity course.It’s a time tracker for Windows and macOS that runs in the background and logs everything you do.An AI layer then sorts that activity into categories like work, leisure, distractions, and personal stuff.
It also breaks down how you spent your day, where you lost focus, and what you could do better.It’s genuinely useful, but two things bothered me.First, paying $120 a year for the basic plan felt steep for what I was getting.
Second, I was sharing a lot of personal information with a third party.Rize could essentially see everything I did on my computer, and I wasn't comfortable with that.So I installed ActivityWatch instead, an open-source time tracker that keeps everything local and private.
However, it doesn't offer any AI-powered analysis.To fill that gap, I took the raw data from ActivityWatch and ran it through an LLM to clean it up, tag it, and categorize it the way I wanted.That data then feeds into an HTML dashboard I vibe-coded to visualize my activity.
Now, you might be thinking that sending all that data to an LLM defeats the point of switching for privacy.It would if I were sending it to Claude or ChatGPT, but I'm not.The analysis runs on Gemma 4 12B, a local model I run through LM Studio, so the data never leaves my computer.
I did use Claude to vibe-code the HTML dashboard, though, since it needed a more capable model.Related 3 things I automate with local AI that I'd never trust ChatGPT with Because your private information deserves a private LLM to process it.Posts 1 By Dibakar Ghosh One $20 AI subscription can save you hundreds You might think I'm replacing a couple of SaaS subscriptions with another AI subscription, but that's not technically true.
For just $20, you can build all three projects I mentioned above and keep using them indefinitely without paying for recurring subscriptions.Any scripts, apps, or websites you create with AI are yours to keep.That's why I believe AI's real value lies in building offline tools for yourself rather than using it casually as a chatbot.
Related Stop building your life around Claude and ChatGPT—make them build something you actually own Don't put all your eggs in one tasklet.Posts 2 By Adam Davidson
Read More
