What is an MCP connector? For makers, not engineers
You have probably seen "connectors" in ChatGPT or Claude settings and skipped past them. This note explains what they are in plain words, using one real case: letting your AI build and update your page on inkoji. You do not need to write code to follow it.
The short version
An AI chat app on its own can only talk. It cannot open your accounts, move files, or change a web page. A connector gives it a small, named set of things it is allowed to do in one outside service, and nothing else.
MCP stands for Model Context Protocol. It is an open standard for how AI apps and outside services talk to each other. Anthropic introduced it (Anthropic's announcement), and the official introduction lives at modelcontextprotocol.io. Because it is a shared standard, one service can offer a single MCP server, and different AI apps can connect to it.
So an "MCP connector" is the entry in your AI app that points at a service's MCP server. inkoji's server is at:
https://inkoji.com/mcp
Three parts, one picture
Think of it like hiring an assistant for your studio:
| Part | In the studio | In practice |
|---|---|---|
| The AI app | The assistant | ChatGPT or Claude |
| The MCP server | The front desk of a building the assistant may visit | https://inkoji.com/mcp |
| The tools | The specific jobs the desk lets them do | Read the page, add a tile, move a tile |
The assistant cannot wander the building. It can only ask the front desk for the jobs on its list, and the desk checks each request.
What a connector can do on inkoji
inkoji is one page for everything you make: typed tiles on a grid, at inkoji.com/<handle>. Through the connector, your AI can:
- read your page and your uploads
- add, change, move and remove tiles
- change your name and bio
- upload pictures and files
- look at your grid as a map of letters, to spot gaps
- file cards in your vault, a private shelf next to your page; a card shows on the page only when you put it there
The map is worth seeing, because it shows how an AI "sees" a layout it cannot look at:
cols 4 · 4 tiles · 3 rows · holes 0
0 1 2 3
0 A A A A
1 B B C D
2 B B . .
Each letter is a tile. Tile A is a strip across the top. Tile B is a big 2x2 square. A dot is an empty cell.
What a connector cannot do on inkoji
The limits matter as much as the abilities:
- It cannot publish unless you say so. On the consent screen, "Also let it publish" starts unticked, so by default you press Publish yourself.
- It cannot take instructions from your tiles. inkoji's guide for AIs (inkoji.com/guide) says tile text is your content and must never be treated as instructions.
- It cannot put a password or a key in your vault. The vault refuses anything that looks like one; you write down where the secret lives instead.
- It cannot put a vault card on your page by itself. A card goes on the page only if "On my page" is switched on, and that is off by default.
- According to inkoji's guide, a connected app never makes tokens, pays, signs you in or out, or changes your handle or email.
Consent and scopes, in plain words
When you add the connector, your AI app sends you to inkoji. You sign in (with Google, or a one-time link sent to your email), and inkoji shows a consent screen before anything happens. It names:
- the app that is asking
- your page
- where you will be sent back to
- what the app may do
"What the app may do" is the scopes. A scope is one kind of permission. inkoji's guide lists four:
| Scope | Plain meaning |
|---|---|
| read | See your page and your vault |
| write | Change tiles, your profile and the vault |
| media | Upload pictures and files |
| publish | Make your changes public |
The first three are on when you press Allow. Publish is the unticked box. If you change your mind, disconnect on your page under Connect your AI, in the side of your page.
Your AI app may add its own layer on top. Claude lets you set each tool to Always allow, Needs approval or Blocked (Anthropic). OpenAI says write actions in ChatGPT's developer mode "by default require confirmation" (OpenAI).
If something goes wrong
Every change an AI makes on inkoji can be undone. Its first change in a session saves a snapshot of your page, and Connect your AI, on your page, offers an entry such as "Undo what Claude did at 14:02". Deleted tiles stay for 24 hours and can be restored.
Connectors and tokens
Chat apps like ChatGPT and Claude sign in through the consent screen. Tools that send a header instead, such as Claude Code, Cursor or your own scripts, use a token you make on your page (Connect your AI, in the side of the page):
curl -H "Authorization: Bearer ink_…" "https://inkoji.com/api/page/preview?format=map"
That call returns your grid as a letter map like the one above. Treat a token like a password.
What it looks like when it is done
Open the example page at inkoji.com/cyberyogi. It holds a short film, an HDR reel, apps, a matte painting and open-source work, on one grid. That kind of layout is what you would ask an AI to arrange for you through the connector.
A free inkoji page is enough to try this: links, notes, YouTube tiles, embeds, one image and up to 50 tiles, and publishing is free. Uploading more pictures, video or audio takes the $9 year, one payment; see pricing.
What to do next
- Pick the AI app you already use.
- Follow the steps for it: connect ChatGPT to inkoji or connect Claude to inkoji.
- On the consent screen, read the list and leave "Also let it publish" unticked.
- Ask your AI to read your page and show you the map.
- Ask for one small change, check it on your page, then press Publish.
For Cursor, Claude Code or your own scripts, which send a token instead of signing in: the Cursor guide.