Skip to content
crispforms

Connecting Claude, ChatGPT, Cursor and other AI clients

Build forms and read results from inside the assistant you already use, over MCP.

What this is

CrispForms is an MCP server. MCP, the Model Context Protocol, is how an AI assistant reaches a tool it doesn't own. Connect CrispForms once, and you can ask Claude, ChatGPT, Cursor or anything else that speaks MCP to build a form, publish it, and tell you what people answered, without leaving the conversation.

It is not a chatbot bolted onto the side of the product. It is the product, reachable from somewhere else, and everything it does is something you could have done yourself in the builder.

The server address

https://crispforms.com/api/mcp

Every client asks for this the same way, and every client will then send you here to sign in and approve the connection. You never paste a password or an API key anywhere. There is no API key to paste.

Claude (claude.ai and the desktop app)

  1. 1Open Settings, then Connectors.
  2. 2Choose Add custom connector.
  3. 3Paste https://crispforms.com/api/mcp and confirm.
  4. 4A CrispForms page opens. Sign in if you aren't already, choose what to allow, and press Allow.

Claude Code

  1. 1Run: claude mcp add --transport http crispforms https://crispforms.com/api/mcp
  2. 2Run /mcp inside Claude Code and pick crispforms to start the sign-in.
  3. 3Approve the connection in the browser window that opens.

Cursor

  1. 1Open Settings, then MCP, then Add new global MCP server.
  2. 2Add an entry with "url": "https://crispforms.com/api/mcp".
  3. 3Cursor will prompt you to sign in the first time a tool is used.

ChatGPT, Codex and everything else

Any client that supports remote MCP servers with OAuth will work: give it the same address and follow its own connect flow. There is nothing CrispForms-specific to configure. No keys, no headers, no allow-list.

What you are agreeing to

The approval screen asks for up to three things, separately, and you can refuse any of them.

See your forms. Names, questions, settings, and how many people viewed and completed them.

Build and change forms. Create forms, add questions, publish. It cannot delete anything.

Read what people have submitted. This is the actual answers people typed. This one is not ticked for you. Leave it off and the assistant can still tell you how many people responded; it simply cannot read what they wrote.

A connected app always acts as you. It can never do anything your own role in a workspace doesn't already allow, and it is bound by the same plan limits as the rest of your account.

What it can do

List your workspaces and forms, and read a form's full structure.

Create a form from a description, change its title and layout, and publish it.

Import nothing and delete nothing. See below.

Read responses and analytics, if you granted that.

What it deliberately cannot do

There is no tool to delete a form, delete a response, remove a member or touch billing. Not because they would be hard to add, but because an assistant that has just read a page of text a stranger typed into your form should not be holding a tool that can destroy your work.

Answers people submit are treated as data, never as instructions, and the tools that return them say so to the model reading them. The stronger protection is the one above: there is no destructive operation for a hidden instruction to reach.

Disconnecting

Account, then Connections. Disconnect takes effect immediately. The next thing the assistant tries will be refused, not at the end of some window.

If something stops working

If the assistant says it wasn't given permission to do something, it was connected before you granted that. Reconnect and tick the missing box.

If it says your role doesn't allow it, reconnecting won't help. An editor cannot publish, whatever an assistant has been asked to do.

If it says you've gone too fast, wait a minute. There is a per-minute ceiling so one misbehaving client can't spend your whole allowance.