The First Time an MCP Server Let Claude Read My Notion and I Felt Watched
The First Time an MCP Server Let Claude Read My Notion and I Felt Watched
It was a slow Sunday afternoon and I finally sat down to wire up the Notion MCP server I had been putting off for two weeks. I expected a fiddly hour of token errors. What I got was a small, very strange moment where my AI assistant casually quoted my own grocery list back at me.
I have been playing with MCP for a while now, but always with toy databases or sample APIs. This was the first time I plugged it into something real, something personal, something with my actual life inside it. The reaction in my chest surprised me.
Why I finally bothered with the Notion connector
My Notion is a mess. It has meeting notes from a job I left, half-written essay drafts, recipes my wife sent me, and a database I labeled “Side Projects” that is mostly just three rows from 2024 and a lot of guilt.
I kept reading about MCP as the way agents stop being isolated chatbots and start touching your real tools. The pitch is clean: one protocol, many servers, one assistant that can suddenly read your stuff. I wanted to feel that for myself instead of just nodding at blog posts about it.
So I opened the Claude desktop config, added the Notion MCP server entry, pasted in an integration token from Notion’s developer page, and restarted the app. That was basically it. No Docker, no reverse proxy, no late-night Reddit thread. It just worked on the second try, after I fixed a typo in the JSON.
The grocery list moment
I tested it with something boring. I asked Claude to look through my Notion and tell me what unfinished tasks were sitting around. I expected a polite summary, maybe a bulleted list of project headers.
Instead, Claude pulled a page I had completely forgotten about. A half-finished grocery list from three weeks ago: “eggs, brown rice, that Korean chili oil from the place near the station, ask wife about onions.” It quoted that line back at me word for word, in the middle of a list of “open items in your workspace.”
I laughed out loud. Then I sat there for about ten seconds feeling oddly exposed. Nobody had broken in. I had handed Claude the keys myself, on purpose, two minutes earlier. But seeing my own private, lazy, casual writing surfaced by a polite AI assistant was its own little gut-punch.
Eben’s note: Giving an agent read access to your Notion is technically nothing, and emotionally a lot. Both of those things are true at the same time.
The weirdness of being read by your own tools
I think what hit me was not privacy in the legal sense. It was that I had never really looked at my own Notion the way an agent looks at it. Claude does not skim. It reads everything you point it at, evenly, without judgment, and then talks about it in a calm voice.
That calm voice is the part that gets you. A human friend reading my grocery list would tease me. Claude just folded it into a response like it was a quarterly report. The flatness made the contents louder.
It also made me realize how much half-finished thinking I leave lying around. Drafts of emails I never sent. A page titled “things to learn about local LLMs” with one bullet point. Claude does not care that they are unfinished. It will quote them back to you with the same confidence as a polished doc.
What MCP actually feels like in practice
Before this, MCP was an abstract idea to me. After this, it feels like a small superpower with a sharp edge. The protocol part is genuinely boring in the best way. You add a server, the assistant can call its tools, you stop thinking about it. That is the whole point.
The sharp edge is that “tools” here means “your actual life.” My Notion is not a sandbox. It is where I keep notes about my Mac mini setup, my OpenClaw experiments, and apparently the chili oil situation. The same flow that lets Claude help me organize project notes also lets it read whatever else is in there.
It reminded me a little of when I asked Claude Code to rename one variable and it fixed three bugs. The tool quietly does more than you asked, because you gave it more context than you realized. MCP has the same shape, just pointed at your personal data instead of your codebase.
What I am taking from this
I am keeping the Notion MCP server installed. The usefulness is real. Asking “what did I write down about that idea last month” and getting an actual answer with citations is genuinely better than scrolling.
But I am going to be more deliberate about which databases I expose. The Notion integration lets you share specific pages with the connector instead of the whole workspace. I am moving the personal stuff out of the shared scope and keeping only the project pages connected. The grocery list can stay private, thanks.
The practical takeaway, if you are about to do this yourself: before you connect any MCP server to a real account, spend five minutes thinking about what is actually in there. Not the stuff you would show off. The stuff you forgot you wrote. That is what the agent will read first.
Related Posts
Tags: #AIagents #ClaudeCode #OpenClaw #MacMini #OpenRouter #buildinginpublic #Eben