As much as my bank hates seeing Apple Music, Spotify, Canva, Adobe Creative Cloud, iCloud+, Google One, Netflix, and about half a dozen other subscriptions coming in every month, I know that AI tools have made an already existing problem significantly worse. Now any AI tool I use wants its cut too. Claude Max, ChatGPT Plus, Google AI Pro, Perplexity Pro, Cursor Pro and more.
Unlike the subscriptions I listed at the beginning, the AI subscriptions are more consistent. Spotify and Canva solve completely different problems. ChatGPT, Claude, Gemini, Perplexity and Cursor, on the other hand, can all do the same things – they’re better at different parts of the job.
It was this overlap that made me wonder how many of them I had to pay for. Instead of looking for one open-source application that can replace everything, I decided to rebuild different parts of my AI workflow with free, open-source alternatives and see how much I could get away with.
January
My local LLM finally feels like ChatGPT (looks good).
After I analyzed the tasks using AI, I realized that a native LLM could easily handle most of my work. Sure, there’s definitely something I’d still like more powerful thinking, bigger context windows, or the special tools that come with boundary models, but most of the things I ask AI to do on a typical day aren’t really required.
While I initially relied on Ollama and LM Studio for all my native AI needs, Ian was a much better fit for what I was trying to replace. Ollama primarily presents itself as a way to run and build with open models, while LM Studio aims to easily run native AI on a PC. Ian’s pitch is more in line with what users like me want. It is directly described as an open source replacement for ChatGPT and Claude.
After using it, I can see why. The biggest difference is that Ian doesn’t make me feel like I’m opening an app just to run a native template. I feel like opening an AI assistant. The interface is very similar to ChatGPT. I can save individual conversations and projects, create custom assistants with individual instructions, load files, switch between different models, and add additional tools through MCP.
Running a model is much less intimidating than the word “local LLM”. Jan has its own model hub connected to Hugging Face, so you can browse models from within the app instead of manually downloading GGUF files and determining if your laptop can actually run them. It’s not just local and is delivered by dozens of remote providers. This means I can use Jan for both my on-premises and cloud models.
What really sells me is that going local doesn’t automatically mean giving up all the extras you’re used to with ChatGPT. Yan has built-in web search and page retrieval, so supported models can still get current information without being stuck with what they’re used to. If you want to go further, MCP allows you to add things like browser control and other external tools!
While Ollama was very good at managing the models, I found that Ian was much better at managing everyone around him. It’s like a middle ground between Ollama, LM Studio and Open WebUI!
Van
Complexity, minus the Pro plan
Vane, formerly known as Perplexica, has long been advertised as an open source alternative to Perplexity, and I can see why! Basically, Vane is an AI-powered search engine that searches the web, collects relevant information, and provides concise answers through attached sources. So, I use Perplexity first.
The difference is that Vane can work with local templates through Ollama, which means I don’t have to send every question to another company’s servers. It uses SearXNG underneath for web searches, and I can choose different search modes depending on what I’m doing. Speed mode is for when I just want a quick answer, Balance mode is for normal searching, and Quality mode digs deeper when you really care about research. I can also narrow my search to things like general websites, discussions, or academic articles, rather than throwing every source type into the same pile.
It sounds pretty simple on paper, but that’s what makes Perplexity so useful to me. I rarely open it because I want to have a long conversation with the AI. I open it because I want to ask questions, get something reasonable back, and immediately see where the information came from.
Vane supports image and video searches, lets you upload files and ask questions about them, can limit searches to specific domains, and even has widgets for things like weather, calculations, and stock prices! That said, Wayne’s opening and running of Perplexity.com is not without friction. You’ll have to go through the trouble of setting it up, and Docker is the recommended route. However, once it’s up and running, I’ll get most of the parts of Perplexity I’m interested in without having to pay for another AI subscription.
Open design
My Second Designer does not require a subscription
While the Claude subscription was once limited to a specific Chat panel, it has since expanded into a full-fledged workspace. You get a relatively new feature called Claude Code, Cowork and Claude Design. One of the best use cases I’ve personally found for AI is using it as an add-on designer.
I’m not a very creative person and spend way more time perfecting slide decks, posters, infographics and even content for social media than I can afford. So I’ve experimented a lot with Claude Design since its launch. He can quickly take a message and turn it into things like slides, one-pagers, prototypes, and other polished visual assets. You can then edit them directly in the interface using prompts or by customizing individual elements yourself. This doesn’t make it feel like a one-shot image generator, and it feels like you have a second designer sitting next to you.
However, after I tried Open Design, an open source alternative based on the same idea, I didn’t feel the need to keep paying $20 to get something like Claude Design. Open design is native and allows you to use the AI agent of your choice as the actual design engine.
This could be Claude Code, Codex, Cursor, Gemini CLI, OpenCode, or others instead of locking you into Anthropic models. It can create things like websites, prototypes, slides, images, and other visual assets, and the output remains as real files that can work instead of sitting on a closed platform.
This flexibility is probably what I like the most. I can keep a general quick workflow for a project I like in Claude Design, but I’m not tied to one template or subscription to get it done. If I want to use OpenCode one day and Gemini CLI the next, I can. I can also implement my own design system via the DESIGN.md file so that the output can maintain the same colors, typography, spacing, and overall style instead of starting from scratch each time.
Open source
Claude Code and Codex, but I prefer the model
For the longest time, the only reason I could justify my Claude subscription was Claude Code. I’ve developed a pretty good system around it, and I use it for everything from fixing small bugs to making changes to entire projects. Although this is very good, I only stick with Anthropic models. Codex is my current favorite, but it comes with the same problem: I can only use OpenAI models with it.
Given that different models have different capabilities, I feel like locking the entire coding workflow to one provider is unnecessarily limiting. There are times when Claude makes better sense of a messy codebase, times when the OpenAI model does the job better, and many smaller jobs where you don’t need to pay for a boundary model at all.
This is where OpenCode comes in. It gives me many of the things that made the Code Claude workflow work, including a dedicated scheduling mode, skills, MCP server support, and the ability to actually read and make changes on a project basis. The main difference is that the agent itself does not depend on the models of the same company. Depending on what I’m doing, I can switch between Claude, OpenAI, Gemini, or even native models without having to rebuild the rest of my setup around them.
Open your notebook
NotebookLM, but I call the shots
As an added bonus, this experience made me realize that I have another AI subscription that probably isn’t worth keeping. I’ve written about NotebookLM far more than any average person, and I still think it’s one of the best AI tools Google has ever created. However, a large part of why I use it is not so much the chatbot itself, but rather the features built around my eyes.
An open notebook comes surprisingly close to repeat that. I can upload my sources, ask questions based on them, make summaries and notes, and even create NotebookLM-style Audio Reviews of the material I’ve contributed. This is one of my favorite features of NotebookLM because I can turn a lot of notes, readings, or research into something I can listen to instead of sitting down and repeating everything.
Of course, NotebookLM is still more polished and has some new features, so Open Notebook isn’t completely duplicated. But if the main things I want are source-based chat, recaps, notes, and Audio Review-style podcasts, Open Notebook covers a surprisingly large part of why I use NotebookLM in the first place.
Related Posts
Microsoft’s rollback of Copilot continues with scaling it out of OneDrive’s file manager options.
This ESP32 CYD project turns your display into a Flock camera scanner
Hackers are using unpatched Plex servers as crypto-miners, and yours could be at risk.
The 3-2-1 backup rule has an obvious caveat that ransomware exploits very well.