The “network effect” commonly refers to a system or service whose usefullness grows with the number of users in the system, often based on the number of connections between users which grows significantly faster than linerally. (A fully connected graph grows at O(n2)). As my use of AI tools (claude in particular) has become more advanced I have found that there is a similar effect in terms of the usefullness of the tools given how much data you can give them access to that follows a similar pattern of usefulness growth.
It would not be wrong to call me a data packrat. I like data. A lot. I scan every meaningful piece of paper that comes into the house and store them; I store location data from my phone; I have just about every email I have sent or received since 1997 (and am bitter I somewhere lost my 1993-1997 college years); I track what books I have and have read; I build wikis for all my D&D games; my home network is instrumented and monitored out the wazoo; my house has many smart devices that are all controlled locally with tons of sensors that save hostorical data. You could say I have a problem; I prefer to say I have a wealth of useful data to be mined.
A lot of the AI coding I have been doing recently has been working to utilize this data via MCP servers so that I can interact with it and make connections between it. A recent example: we need to have our roof redone. I know that the last time was 10-15 years ago and I wanted to find out how much it had cost last time. I thought I had that information in my email but I couldn’t find it. So, I asked claude to search my email and find it. It was able to find some related emails, but not the data that I needed in email. However, using the information in my emails, it searched my PDF archive and found where I had scanned the invoice and was able to get my answer. Like any assistant, the more data they have access to, the better they can do their job.
So what have I been building:
- Arkivo - Giving AI tools access to my email was something that took me a while to do. I had many concerns. Sending email on my behalf was not something I wanted, but that is simple to prevent. My bigger concern was that an AI tool could use even read-only access to my email to gain access to services I use by triggering password reset emails and then reading them and using the links. So, I built Arkivo. It imports email from Gmail, Fastmail, or any IMAP server and makes it available via an MCP interface with powerful search capabilities. But with one key restriction. It allows the owner (me) to set a waiting period before emails become indexed and searchable. I set mine to a few days. What this means is that any password or account reset email will have expired before my AI tools get access to it.
- Lokilo - For years I have used Owntracks to record my locations as I move around. All my vacations and travels around the world I have recorded. I love the system but their backend service was not the most advanced. But since they are an open project, I was able to build a complete replacement backend that receives data from the official iOS client, stores it in a more searchable manner, and exposes a nice web interface, and an MCP interface.
- Advanced Memory MCP - I wanted all my AI tools to have a shared knowledge graph they could use to store information for later. Think of it as long term memory. (As the tooling has advanced, claude and other tools have added this internally, but I prefer to own my data so I can share it across tools.) I instruct claude via
CLAUDE.mdto store useful bits of information whenever we talk and it does. - Hardcover MCP - This allows me to access all my books and reading data from Hardcover.
- Paperless MCP - I extended an open source MCP with even more API coverage to get full access to every PDF I have ever scanned.
- Homeassistant MCP - I can control items in my house and get access to sensor data. This has allowed me to analyze trends in heating and cooling and how it impacts power consumption and electricity usage.
- Zabbix MCP - Another open source MCP I expanded to have even more functionality. I am able to monitor, diagnose, and often solve network and computer issues because of the data that is accessable.
This isn’t even all of the ones that I use. I use MCPs to access GitHub, DnD Beyond, my private Wikis, and my calendars. But this was a lot and managing them all in claude was becoming a problem so there is one final piece that brought this all together. If you use your own MCP servers, you will find after a while that managing them becomes very difficult. Some will only run locally on your computer, making them only available on your computer and not your mobile devices. Some are network based but you need to share them publicly on the internet to have them avaiable and their security may not be something you are comfortable with. And setting them up is just a pain.
So I built an aggregator:
- MCP Hub - This allows me to host a single, internet accessable server with strong security that either runs CLI MCPs internal to itself or proxies access to other network based MCPs that can be run in isolated environments not accessable to the wider internet. It allows me to connect claude to it across all platforms, desktop and mobile, so I can use the same set of tools from everywhere.
So this is my network. A network of data that has made using AI to help me far more useful. In my opinion, this is the only way to integrate AI tools into your workflows. I don’t want every company slapping AI onto every product in isolation. That just forces me to make the connections between things. I want companies to allow me to access my data via APIs and MCPs so that I can build my own personal web of data and garner the network effect benefits myself.