MCP

The MCP server is the same surface spoken in a conversation. Connect Claude or ChatGPT, tell it what happened, and it captures, reads back and reviews through the same tools the API exposes. This page shows how to connect and what to expect.

Over MCP you do not post JSON. You tell an assistant about your day and it calls the same operations the API has, with the same rules and the same answers. The MCP domain speaks Streamable HTTP at its root; the examples use mcp.undrstand.com.

MCP URL

https://mcp.undrstand.com

Signing in creates your account if you do not have one. Claude and ChatGPT count as you, so there is no consent screen: they get the whole surface.

Connect Claude

In Claude, open Settings, then Connectors, and choose Add custom connector.
Paste the MCP URL as the server URL. Leave the OAuth fields empty; Claude discovers them.
Claude opens the connect page. Enter your email for a sign-in link, or use a social account, and you are back in Claude.
Turn the connector on in a chat and start talking.

Connect ChatGPT

In ChatGPT, open Settings, then Connectors, and turn on developer mode under Advanced.
Create a connector with the MCP URL and OAuth as authentication. Leave client id and secret empty.
Sign in on the connect page, then enable the connector in a chat.

Connect Claude Code

In a terminal, add the server and authenticate once:

claude mcp add --transport http undrstand https://mcp.undrstand.com
claude            # then type /mcp and pick undrstand to sign in

The tools

Every tool is a REST route spoken differently. The assistant picks them; you do not have to name them.

Tool	Does
`add_capture`	Hands in your words exactly as you said them.
`view_captures, view_capture`	Lists what you handed in, or reads one back with its segments and tags.
`delete_capture`	Removes a capture and what was found in it.
`view_patterns, view_pattern`	Your patterns, or one in full with its captures, correlations and flows.
`review_pattern, review_correlation`	Records that a pattern or correlation is yours, or is not. Only when you say so.
`view_timeline`	Your days between two dates, with captures and patterns.
`view_settings, update_setting`	Reads or changes your timezone and language.

How a conversation goes

Tell it what happened. Not a summary, not a mood score: the situation, what it did to you, what you did next.

My boss called at six again about the report. I felt that knot in my stomach and spent the evening scrolling instead of finishing it.

The assistant hands it in as you said it, in your language, and tells you it is being processed. Ask it for a read-back later, or just carry on with your day.

After a few days, ask what it sees.

What patterns do you see so far?

It reads your patterns and tells you in its own words: what keeps setting you off, which feelings come with it, what you tend to do, and who or what it happens more with. Ask for one in full and it tells the flows: what happened, what you felt, what you did.

When something is off, say so. Only then does it record a review.

The procrastination one is right. The one about my sister is not a pattern, that was one bad week.

Privacy

What you tell the assistant goes to undrstand and is stored on your account. The assistant's own memory is a separate matter, governed by Claude or OpenAI.
The assistant sees your data through the tools, never anyone else's.
Disconnect the connector in Claude or ChatGPT to revoke its token. Delete your account to remove everything.
The assistant is told not to diagnose, not to label you, and not to name the method behind the analysis.