This is the swap, told honestly. We did not theorize about whether open-source local tools could stand in for the paid ones we were using. We made the replacements, one category at a time, and ran the studio on them. Some swaps were boring in the best way. One asks something real of you on the first setup. Here is the whole mapping and what each trade actually felt like.
We are not going to hand you a savings figure or a tidy count of “tools eliminated.” Those numbers depend on which subscriptions you happened to carry and what you actually use. What we can give you is the method we used to count and the category-by-category replacement that came out of it.
How we decided what to replace
We did not start with the tools. We started with the work. We listed the jobs the studio does in a normal week that touch AI: everyday chat and drafting, coding, transcription for documentary work, asking questions of our own documents. For each job, we wrote down what we were paying for and asked one question: is there an open-source tool that does this job locally, well enough for real work, on the machines we already own.
If you want to run the same count for yourself, the method is the part that travels, not our number. Open your billing. List every AI subscription. Next to each, write the actual job it does for you, not the marketing. Then check each job against a local alternative. The count of tools you can replace is yours to find [CONFIRM: exact number of paid tools the studio replaced].
The one engine underneath everything
Before the mapping, the piece that makes the mapping possible. Almost every paid AI tool is a nice interface sitting on top of a model someone else hosts. Locally, you separate those two things. You run one model engine, and you point many tools at it.
For us that engine is Ollama. It pulls a model with one command and serves it on http://localhost:11434, speaking the same API shape the big cloud providers do. That last part is why this whole article works: a local tool that “speaks ChatGPT” can usually speak to Ollama instead, with no code change. One engine, one endpoint, many tools. We pulled the model once and every replacement below reuses it.
Paid chatbot to Open WebUI plus Ollama
The everyday cloud chatbot, the one you open to draft an email or think out loud, was the first swap and the easiest to live with.
Open WebUI is a browser-based chat that looks and behaves like the paid product people already know: conversation history in a sidebar, model switching from a dropdown, file uploads, system prompts. It runs locally and points at our Ollama model. The honest cost of the swap is the install: Docker is the smooth path, and the first run has you create an admin account. After that, it is a chat app that happens to never send your words anywhere. Open WebUI even has built-in document upload, which means light “chat with this PDF” work does not need a separate tool at all.
Cursor to OpenCode plus Ollama
The paid coding agent was the swap we cared most about, because code is the thing you least want leaving the building.
We replaced it with OpenCode pointed at the same local model. OpenCode is unusual among open-source coding agents because it ships as a normal desktop app: you open a window, not a terminal. It reads files, proposes edits, runs commands, and asks permission before it acts. Wired to Ollama, no code goes to a server. We covered the exact wiring in a separate build log, including the minute it confused us, but the headline for this article is that the paid agent’s shape carried over without a terminal and without a subscription. The one honest caveat is the same for every local coding setup: the model is the ceiling. On genuinely hard problems a frontier cloud model is still ahead, and OpenCode lets you switch to one for that single task and switch back.
Paid transcription to WhisperX
This is the swap with the sharpest before-and-after for our documentary work, and also the one that asks the most of you up front.
We were choosing between paid transcription services for diarized output, the kind that labels who said what, or plain Whisper that gave great words in a format useless for editing multi-speaker recordings. WhisperX gave us both: transcription plus speaker diarization plus word-level timestamps, running locally. Per the Atlas note, the diarization quality is excellent for two-to-four-speaker recordings and gets shakier with five or more talking over each other. The thing it asks of you is a one-time HuggingFace token and license acceptance for the diarization models. It is annoying once and trivial forever after. On a Mac CPU a long interview takes meaningfully longer than on a GPU, so this is a “start it and walk away” tool, not an instant one.
Cloud document Q&A to AnythingLLM
The last swap covers “chat with our documents,” which for a studio means client briefs, reports, and interview transcripts.
AnythingLLM is a polished chat-with-your-documents app: drag in a PDF, ask questions, get answers with source citations you can click back to. It runs as a desktop app for single use or a Docker server for a small team, and it points at the same Ollama model. The citations are the reason we trust it over a generic chatbot for this job, because you can verify the answer against the original passage. The honest trade is that the default chunking and retrieval are good-enough rather than state-of-the-art, and complex PDFs with tables or multi-column layouts benefit from cleanup before ingestion. For the everyday “ask our own files” job, it replaced the cloud version cleanly, and nothing we uploaded left the laptop.
What the swap actually cost
Nothing leaves the building now, and there is no monthly line that grows as we add tools. But honesty matters more than the win, so here is the real cost. You trade a polished hosted experience for a setup you own and maintain. Open WebUI wants Docker. WhisperX wants a token the first time. The coding agent is only as strong as the local model you give it. None of these is hard, but all of them are yours to keep running.
What we got in return is worth that cost for us: the work keeps running on a plane and during an outage, the privacy story is observable rather than promised, and the whole stack is reproducible from open parts. We did not replace our paid tools to prove a point. We replaced them because the local versions did the jobs we actually do.
Curious about these things. You should be too.
Harness your curiosity.
— Stridenote · № 003