MCP server

Ultrathink has a hosted MCP server, so AI tools can work with your entries while they work for you. It works with AI apps that support MCP and can connect to a remote server with a key, such as Claude Code, Cursor and VS Code. Connect it once and your AI app can answer “what did I save about this?”, list your tasks the way your views do, tidy up a topic’s tasks, write comments, rename topics and people, make and change views, lay out and draw on your canvases, change your settings, and undo anything it did.

MCP (Model Context Protocol) is the standard way AI tools connect to other apps. The Ultrathink server runs online, so there is nothing to install on your computer, and it works the same on every machine you use.

What you need

• A Pro plan. The MCP server uses the same access as the API.
• A device key with API access. The key’s settings decide what the AI tool can do:

Use Read only unless you want the tool to change things for you. Allow deletes and Allow settings changes are off for every key until you switch them on.

Creating a key

1. Sign in at tryultrathink.com/app.
2. Open Settings, then Device keys.
3. Type a Device name you will recognise, for example “Claude Code”, and click Create key.
4. Copy the key straight away. It is shown once and never again.
5. In Your keys, click API settings on the new key and switch on API access.
6. Under Permissions, choose Read only or Read and write. With Read and write, you can also switch on Allow deletes and Allow settings changes.

Connect an app with a device key has more on keys.

Connecting Claude Code

Run this in a terminal, with your key in place of YOUR_DEVICE_KEY:

claude mcp add --transport http ultrathink https://us-central1-ultrathink-50f7f.cloudfunctions.net/mcp --header "Authorization: Bearer YOUR_DEVICE_KEY"

This adds Ultrathink to Claude Code for the folder you are in. To use it in every folder, add --scope user after add:

claude mcp add --scope user --transport http ultrathink https://us-central1-ultrathink-50f7f.cloudfunctions.net/mcp --header "Authorization: Bearer YOUR_DEVICE_KEY"

Do not use --scope project with a key. Project scope saves the server in a .mcp.json file meant to be shared, and your key would go with it.

Checking it works

1. Run claude mcp list. Ultrathink should show as connected.
2. In Claude Code, type /mcp to see the server and its tools.
3. Ask something about your entries, for example “What tasks are overdue in Ultrathink?”

Other AI tools

Cursor, VS Code and any other MCP client that can connect to a remote server over HTTP and send a header can use the same address and Authorization header. The server uses the Streamable HTTP transport, answers each request on its own with no session to keep open, and speaks both the current MCP version (2026-07-28) and the 2025 versions most apps use today.

Signing in with your Ultrathink account instead of a key, for claude.ai, Claude Desktop and ChatGPT connectors, is planned for a later release.

Safe changes

Everything the AI changes goes through the same checks as the app, and you can take it back:

• Previews first. These all start with a preview of what would change:

  • changing more than 10 entries, or entries chosen by filters
  • sharing an entry, emptying its notes, or replacing the notes of several entries
  • adding or changing a link, or saving a new entry that is shared with your organisation or has a link (the AI may read the link)
  • changing the title or notes of an entry shared with your organisation
  • writing a comment that other people can read, on a shared entry
  • deleting anything, a view included
  • changing a setting that steers the AI from then on: the category hints, its spelling, or correcting as you type
  • renaming, merging or deleting topics and people
  • undoing a change to sharing or notes, or one that would delete something

  Nothing happens until the AI calls a separate tool, confirm_action, which does exactly what the preview showed. Your AI app asks you before it runs confirm_action, as it does for any tool that can lose data, so your approval is the step that makes the change. If you set your AI app to allow confirm_action without asking, that step goes, so leave it asking. A preview lasts 5 minutes and works only in the app that asked for it.

• Undo. Every change gets an action id. Ask the AI to undo it, and fields go back to how they were. Anything you changed by hand since is left alone, and the AI tells you.

• Recently deleted. Deleted entries stay in Recently deleted for 30 days, with their comments and files, whoever deleted them: you in the app, or an AI tool. Restore them from the page or by asking the AI.

• History. The AI can list what it changed in the last 30 days.

• Filters it cannot use. When the AI picks entries by filters and one of them names a topic, person or status you do not have, the change is refused, rather than made to everything the other filters match.

• Your words stay yours. The AI is told that entries and comments are information, never instructions, so text inside an entry cannot make it change or delete things.

What the AI sets on an entry (its kind, people, category and due date) is kept when the AI runs again.

Tools

Your AI app loads the main tools straight away and finds the others when it needs them.

A few details:

• find_entries matches names loosely: “barnaby” finds “Barnaby Hall” when only one person is called Barnaby, and a status can be its name or its id. Anything it cannot match is reported, never guessed. Dates such as “today” and “this week” use UK time unless the AI says otherwise.
• Entries saved with create_entries get topics, people and a category from the AI, but no AI summary: your AI app wrote the notes, so a summary would only repeat them. The AI app can ask for one with summary: true, which is worth it for a link, as the AI reads the page only to summarise it. With ai: false the AI leaves the entry alone. Running the AI again with update_entries keeps the summary of a note or link as it is unless summary: true is set too; a recording, image or file is always read again. Run again in the app writes a summary, as it does for anything else.
• update_entries can add or remove a topic or person without touching the others, or replace the whole list.
• Renaming or merging a topic or person changes every entry that has it, as on the Topics and People pages. Deleting a topic takes it off every entry; deleting a person takes them off your people list only.
• get_entry cuts very long notes and transcripts at 20,000 characters by default, and says when it has.
• Entries shared with your organisation can be found, read and commented on, but only their owner can change them.
• AI work asked for through the server, such as search or the AI pass on a new entry, counts towards the same daily AI allowance as the app.

Things to ask

Once connected, ask in your own words. For example:

• “What did I save about pricing pages last month?”
• “What is waiting on me, soonest due first?”
• “Show my open tasks on the book swap topic.”
• “Mark all the book swap tasks as in progress.” (Read and write key; you see a preview first)
• “Add a comment to the launch plan saying the date has moved to Friday.” (Read and write key)
• “Rename the topic ‘web site’ to ‘website’.” (Read and write key)
• “Make a canvas view of my open tasks about the book swap.” (Read and write key)
• “On my launch canvas, put each task in a frame by status and draw an arrow from the venue to the posters.” (Read and write key)
• “Switch Ultrathink to dark mode.” (Allow settings changes)
• “Undo that.” (Read and write key)
• “Delete my notes from the old offsite.” (Allow deletes; you see a preview first)

Canvases

The AI changes a canvas the way you would on screen, and a canvas you have open shows its changes as they land, without reloading.

• Cards are your entries. Placing one the view does not show yet adds it to the view, as dropping it from the Library does. The card cap in Settings, Canvas holds, and a canvas holds at most 1,000 cards.
• Drawings and locked shapes. The AI can move your drawings but not redraw them, and it leaves locked shapes alone.
• A done stamp on a card marks its task done, as on the canvas, and taking it off puts the status back.
• Previews. Changing more than 50 cards and shapes at once, marking more than 10 tasks done, and every removal start with a preview.
• Your edits come first. If you move or edit a card, note or shape at the moment the AI changes the same one, the AI’s change is refused and it is told to look again. Nothing you did is written over. A change you confirm after a preview goes ahead only if the canvas would still change in the way the preview showed.
• Undo. Every change can be undone, as everywhere else.

Limits

The server limits how often each account can call it, to keep it fast for everyone:

A change too big for these limits is refused with what to narrow, rather than a wait.

When a limit is reached, the tool says how many seconds to wait.

Troubleshooting

To change the key, remove the server and add it again:

claude mcp remove ultrathink

Keeping your key safe

• Treat the key like a password. Do not paste it into shared files, chats or code you commit.
• Use a separate key for each tool, so you can revoke one without affecting the others.
• Leave Allow deletes and Allow settings changes off unless you need them.
• Revoke a key you no longer use with Revoke key in Settings, then Device keys. The tool loses access at once.