Connect Claude.ai to your notes with a custom connector

Most of what I ask Claude for in the browser is something I will want again: a summary of a long thread, the draft of an email, a comparison of two options, the plan for a weekend project that will take three weekends. Until recently the only way to keep it was to copy it out of the chat into a notes app, and the only way to send it to someone was to paste it again somewhere else. Two copies, one of them already out of date.

A custom connector removes both steps. Add one address in Claude.ai's settings, sign in once, and Claude can write into your notes, search what is already there, and put a page on the web to hand you a link. No terminal, no token, nothing to install. It looks like you are trying to save a note. Would you like help with that? Yes. For once, yes.

This post is the setup, what you can then say to Claude, and exactly what the connection is allowed to do.

🔌 What a custom connector is

Claude.ai can talk to outside tools over an open standard called the Model Context Protocol. A tool that speaks it and lives on the web is a remote MCP server, and Claude.ai's name for one you add yourself is a custom connector. Anthropic's help centre lists them as available on every plan, the free one included, in Claude.ai and Claude Desktop. The free plan allows one connector, so choose wisely, or at least choose this one.

The connector for Sumibako is one address:

🛠️ Setting it up

1. Open Claude.ai's settings and find Connectors. On Pro and Max it sits under Customize; on Team and Enterprise it is in the organisation settings, and an owner has to add it. The help centre article above has the current screenshots, because menus move and blog posts do not.
2. Choose Add custom connector, paste the address into the URL field, and save.
3. Press Connect. You are sent to sumibako.com to sign in, or to create an account if you do not have one. Claude never sees your password. It is an ordinary sign-in on our site, the same shape as "Log in with Google" anywhere else.
4. Read the consent screen. It says what the connector may do: read and search every page in your vault, write new pages, and publish a page it wrote so it can hand you a link. Approve it.

That is the whole setup. Back in Claude, the tools appear on their own, and nobody pasted a token into a chat window, which is the part I am proudest of and the part nobody will notice.

🗣️ What to say to it

You do not call the tools by name. You ask for the outcome and Claude picks the tool.

• "Write this up and save it to my notes." Claude files a page, with an icon it chose, under a page called Agent notes in your vault.
• "Find my notes on the pricing decision." Claude searches your vault by title and body and reads the page back.
• "Add a line to my decisions page saying we chose Postgres." Claude appends to the page without rewriting the rest of it.
• "Publish that and give me the link." Claude puts the page on the web and returns the address. Anyone with the link can read it. Search engines are told not to index it.
• "What was the plan I saved on Tuesday?" Claude searches, reads, and tells you, which is more than Tuesday's version of you would have managed.

Pages Claude writes are ordinary pages. Open them in the app, edit them by hand, move them, change the icon if its sense of humour is not yours, and the connector sees the change the next time you ask.

🚫 What it cannot do

I wrote the connector, so I want to be precise about its limits, because the consent screen is where most people stop reading, and I would rather you stopped here instead.

• It cannot delete anything. There is no delete in the connector at all. Taking a page out of your vault is something you do in the app.
• It cannot publish a page you wrote yourself. It may put on the web only pages it created. If you write something private in the app, no prompt to Claude can make it public.
• It cannot make a page findable by a search engine. A published page is a link and nothing more, unless you turn on indexing for that page yourself.
• It cannot reach any account except the one that approved it.
• It cannot see your password. The sign-in happens on our site and hands Claude a token that is limited to reading and writing pages.

To disconnect, open Settings in Sumibako, then Connected apps, and revoke it. It stops on its next request. Removing it from Claude's connector settings works from the other side.

🐢 The part that took the longest

Making the connection work with nothing to paste took longer than the connector itself. Claude.ai's connector form is built around OAuth, the same sign-in dance that "Log in with Google" uses. That means our server has to tell Claude where to sign in, Claude has to register itself with that sign-in, and the token it gets back has to say what it is for and be checked on every request. Getting each of those right against the current version of the standard, rather than against a memory of an older one, is most of why there was a gap between "the terminal works" and "the browser works". The standard moved while I was reading it, which I am told is normal.

I mention it because the result is invisible. You paste an address and press Connect. Everything that makes that safe is behind the button, sulking.

🤖 ChatGPT

ChatGPT has a connector setting of its own, on some plans, and the same address goes into the same kind of form. The flow is meant to be identical: it asks our server how to sign in and sends you to the same page. I have run the Claude.ai flow end to end. At the time of writing I have not yet done the same for ChatGPT, so treat that one as expected rather than confirmed, and tell me if you get there first.

🤷 Why bother

Because the thing you asked Claude for is worth keeping, and a chat is the worst place to keep it. It scrolls away, it cannot be searched properly, and it cannot be sent. A page in your notes with a link can be all three. The connector is just the shortest path from the answer to the page. The paperclip would be proud.