I was paying for four AI tools at once, and each one covered a job the others handled rather poorly.
ChatGPT answered my daily questions and is excellent for images and infographics. Claude handled research, brainstorming, and my occasional coding endeavors.
NotebookLM indexed and summarized my own documents, and Perplexity handled anything that needed live sources because I don’t trust any of the others for proper fact-checking.
As you can imagine, managing these subscriptions can become complicated. That’s not including the one-off token purchases for when I exceed my usage limits.
After a few months, the total stopped making sense. So I went looking for open source replacements, one for each job, running on hardware I own.
Open WebUI gave me a mobile chat window
One address for every model, with no monthly bill
Open WebUI runs in Docker on my computer and gives me a browser interface that behaves like the ChatGPT web app. It connects to Ollama for models running locally or to any OpenAI-compatible endpoint if I want something bigger.
The most important part for me is that it works properly in Chrome on Android. I added it to my home screen, and it opens like an app, with my full chat history already loaded.
I can switch models in the middle of a conversation, and no paid app lets me do that. I can also drop documents into a knowledge collection and attach that collection to a chat, so the model answers questions based on my files instead of guessing.
The one downside is that its license requires you to keep the Open WebUI branding in place. Personal and home use is unaffected, but this requirement is why some people argue it no longer counts as open source.
PocketPal AI runs directly on my phone
On-device models with no account or connection
This is the app that replaced Claude for me. PocketPal AI is an MIT-licensed Android app on the Play Store that downloads models from Hugging Face and runs them on the phone.
No data leaves the device. No sign-in, server, or data allowance is spent, and this matters more to me than someone on an unmetered connection.
A phone can only run a small model, so I run something in the 1B to 4B range. That is enough for throwing ideas back and forth for my coding projects, generating boilerplate code, or working through an idea while I am away from my desk.
It includes a benchmark tool that measures tokens per second on your device, which is a fast way to find out what your phone can handle before you use up storage for a download.
I also set custom system prompts, so the model starts each session already knowing how I want it to write.
It has performance tradeoffs on both hardware and the model. Continuous generation warms the phone and drains the battery, and a small model tends to drift on anything complicated.
AnythingLLM handles my private documents
Workspaces that keep each project’s sources apart
AnythingLLM is an MIT-licensed app from Mintplex Labs that turns a bunch of files into something I can question. I upload PDFs, Word files, plain notes, or web pages, and it indexes them without any setup.
Everything runs in workspaces that stay isolated from one another. My work research and personal notes are never mixed up, which is something NotebookLM handles poorly.
Answers come back with citations pointing at the source document, so I can check a claim without reopening the original file.
The desktop version installs in one click and ships with its own model engine. I connected mine to Ollama instead, which lets me run a larger model than the bundled one.
There is also an Android app that pairs with my instance over the network, so I can access my workspaces anywhere.
Vane answers web questions with sources I can click
Live search that is not tied to one company’s index
Vane is the project formerly called Perplexica, renamed by its maintainer in March 2026. It ships as a single Docker image that bundles its own SearXNG metasearch engine, so there is very little to configure.
It searches the live web, then writes an answer with numbered citations. I click through the citations more often than I did with Perplexity, mostly because the sources are easier to access, and the habit is easy to cultivate.
There are three modes for speed, balance, and quality, so I can pay for depth only on questions that need it. I can also direct it to the web, discussion forums, or academic sources, depending on what I am researching.
On my phone, I added it as a custom search engine in Chrome. Typing a question in the address bar now sends it to my computer instead of Google, which is a smaller version of what happens when you run a local LLM on Android for offline answers.
Open source still has a long way to go
Anything that needs heavy reasoning still goes to a paid model. Long analysis, tricky code, and multistep tasks quickly reach the limits of these free tools.
Speed is the other limit. A Raspberry Pi hosts these interfaces without issues, but it generates text slowly, so the heavy lifting happens on a machine with a proper GPU or through an API key.
Setup isn’t difficult, though it isn’t nothing either. Each tool is a Docker command and a config screen, which is about the same effort as replacing paid subscriptions with open source apps in any other category.
What I gained in return was data ownership and privacy, offline access, and fewer subscriptions to manage.
