What the experiment replaced
The author's setup, as reported:
- Search: a local language model (Qwen 3.5 at 4B or 9B parameters, chosen by how much GPU memory was free) with web access through Brave Search in LM Studio and a web tool in Jan.
- Docs: OpenKnowledge, a markdown editor that opens an Obsidian vault.
- Photos: digiKam, a desktop photo manager, chosen over self-hosted options like Immich and PhotoPrism because the author had set those up before and didn't want to repeat it for a one-week test.
The week's results are the author's own experience, not an independent test. The same goes for the judgment that this is "smarter" than deleting the account.
Search: a local model doesn't mean local search
This is the part that's easiest to misread. A model running on your PC knows only what it was trained on, so current facts need web access. The author says their queries now go to Brave or Exa instead of an account that also holds their email. That is a change of provider, not a move to offline search. Your queries still leave the machine.
Jan's documentation shows how this works today:
- Jan v0.8.4 added built-in web_search and web_fetch tools, so models can search and fetch page content without an external MCP server.
- The tools are offered to any model that supports tool use. That is why the author favors Qwen models for their tool-calling.
- Web search is on by default, and the default provider, Exa, uses a free keyless endpoint.
- The setting lives under Settings > Web Search. Jan's docs also list Tavily, SearXNG and You.com as providers.
- Jan's docs say SearXNG is self-hosted, and Tavily needs an API key. Self-hosting SearXNG is the route for people who want to cut third-party search exposure further.
The author mentions a "Web MCP in Jan". Jan's docs say earlier versions relied on an Exa or Serper MCP server for web search. Current builds don't need one, so you may not need to set up an MCP server at all. Jan's quickstart still says you can configure the Exa MCP server under Settings > MCP Servers with your own API key if you prefer a different provider.
Brave's route is separate. The official Brave Search MCP server package lists a Brave API key as a required environment variable. That means you need a key from Brave, which adds setup the story doesn't describe.
Jan's own agent documentation gives advice that applies to any local-model search setup. Web results are only as current as what the search returns, and a model will still summarize what it reads, so for anything important you should follow the link. The author's trade-offs fit this. Maps has no equivalent here, and web-augmented answers use more tokens. In the author's view that pushes you to write more careful prompts and read the sources yourself.
Documents: markdown, Git and AI
OpenKnowledge describes itself as a three-layer system. It has a WYSIWYG editor for you and MCP and skills tooling for agents, and your content stays as plain markdown or MDX files in your project, version-controlled with Git. Its docs list a Windows desktop app alongside macOS 13 or later and Linux, plus a local web app and CLI.
That matches the author's report that history and undo run on Git without any Git commands. Nothing the sources show says it matches Google Docs' real-time multi-person collaboration or version-history behavior. Treat it as a different tool that covers the writing experience, not a clone.
"The files don't leave my drive" needs qualification. OpenKnowledge's privacy documentation says content stays on the machine by default. It also lists features that can send data off it, and those are opt-in or on-demand:
- semantic search
- diagnostic or bug-report bundles
- GitHub sync
It also lists two automatic exceptions that don't touch your documents. The desktop app checks for updates, and on first launch it opens your browser once to check for a pending share link.
The AI layer is where privacy depends on your choices. The author says Claude edits appear live in the document, and that you can connect many providers, including a local LLM. If you connect a hosted provider, document text may go to that provider. If you want the editor to stay private, point it at a local model. The story doesn't say which provider handled the author's edits.
Photos: a library manager, not a backup service
digiKam is a free desktop app that indexes folders where your photos already live. The author says it doesn't upload or copy anything. Per digiKam's feature list:
- Its deep-learning model files are optional, downloaded on demand, and run on your computer with no cloud data.
- The AI features are built on OpenCV and can be GPU-accelerated.
- The AI features include face detection and recognition, auto-tagging, aesthetic detection and auto-rotate.
- Duplicate finding and an editor with image versioning (so originals aren't lost) are also on the list.
The author reports downloading about 1.2 GB of models and restarting before face recognition and auto-tagging worked. That figure is firsthand. The official pages I reviewed don't confirm the size. They do confirm the pattern: digiKam's manual says face management needs access to deep-learning models that may be downloaded from the web during first run. So the first run needs internet access even though the processing is local afterward.
Windows requirements and warnings
digiKam's manual lists the following for Windows:
- It runs only on 64-bit Windows 10 or newer, and is not guaranteed on server or embedded Windows.
- The recommended RAM is 16 GB for AI features, with 8 CPU cores suggested for AI use.
- It recommends a GPU with OpenCL support (such as NVIDIA cards) to speed up the AI engine.
- There are two packages. The installer needs administrator rights, and the standalone archive doesn't.
- The installer is unsigned, so expect a Windows security warning and a Defender dialog on a second run.
In managed environments, that unsigned installer will likely need an exception. The standalone archive may be easier for a non-admin user. For domain users, the manual says all users need read/write access to digiKam's AppData folders.
Face-data caution
The manual warns that "Reset and clear all faces and training" deletes all face training and rescans every image. Confirmed faces go back to Unknown, and you retrain from scratch. If you tag a lot of faces, back up the databases first. "Rebuild all training data" is the gentler option.
The limit that matters
The author is clear that digiKam has no phone backup, which is probably why many people use Google Photos in the first place. Getting photos onto the PC is manual. This setup organizes a library you already have. It doesn't capture one. If your phone is the camera, you still need your own transfer routine, and you need a backup plan for the library too. Local storage is only as safe as your drives.
What this means for Windows users
- A partial move works. The author's conclusion is that the Google account no longer holds everything, and that this beats outright deletion. Nothing here contradicts that.
- Know the boundary of "local". Search traffic, hosted AI providers, first-run model downloads and optional sync are all places data can leave the PC.
- Hardware matters. Local models and digiKam's AI features lean on RAM, CPU and GPU. A modest laptop will feel it.
- Maps, email and phone photo backup are the gaps. None of these tools fill them.
A sensible way to try it:
- Install digiKam (standalone archive if you lack admin rights) and point it at an existing photo folder.
- Let it download the face and tagging models on first run, then restart.
- Install Jan, pick a tool-capable local model, and leave built-in web search on, or switch the provider to SearXNG if you self-host.
- Try OpenKnowledge on a copy of a notes folder before trusting it with the originals.
- Decide per tool whether any hosted AI provider is acceptable.
The experiment isn't a blueprint for leaving Google. It shows which parts of three workflows can be moved to your own PC, and where the limits are.