Let your AI assistant read this workspace

Create an access token

1. 1

   Open Account & Team in the sidebar.

2. 2

   Choose AI assistants.

3. 3

   Type a name into Name.

   Name it for the assistant and the person, so a list of them is readable a year later. The placeholder is "Claude on my laptop". Sixty characters at most.

4. 4

   Choose a period under Stops working after.

   • 30 days — for a one-off piece of work.
   • 90 days (recommended) — the default.
   • One year.
   • No expiry — it keeps working until somebody revokes it. The screen warns you to choose this only for something you will remember to switch off.
5. 5

   Tick Also let it propose changes only if you want it to.

   Leave it off for an assistant that should answer questions. What it allows is set out further down this page.

6. 6

   Click Create token.

7. 7

   Copy the token from the yellow panel straight away.

   It begins with `mq_`. The panel says it will not be shown again, because it will not.

8. 8

   Click I have copied it to dismiss the panel.

Connect your assistant

Any assistant that speaks the Model Context Protocol over HTTP can connect, and it needs two things: the server address and the token.

1. 1

   Copy the server address from the Connect your assistant panel.

   It is your own MarqIQ address followed by `/api/mcp`. Copy it from the screen rather than typing it, so it matches the address you actually sign in at.

2. 2

   Choose your assistant from the tabs under Step by step, for the assistant you use.

   There are tabs for Claude Code, Claude Desktop, Cursor and VS Code (Copilot), each carrying the exact line or block of configuration that client expects.

3. 3

   Copy the configuration with the Copy button on the code block.

4. 4

   Replace YOUR_TOKEN in it with the token you copied.

5. 5

   Follow the numbered steps under the code block for that client.

   They say where the file goes and how to check the connection came up.

6. 6

   Ask it something.

   "What should I fix first in HubSpot?" or "Run the weekly review". It says which period its numbers cover, and it says so when MarqIQ does not have enough data to judge. When the audit behind an answer is not ready, the assistant is told what you would be told — “MarqIQ is still working out the data audit for HubSpot; this answer needs it”, with how far it has got — and can ask again later.

Whatever the client calls the field, the token travels as a header named `Authorization` whose value is the word `Bearer`, a space, and the token. In an assistant's settings that is usually a field called "API key", "bearer token" or "custom header" under add a remote MCP server or add a connector. A token is never read from a web address, because addresses end up in logs and histories.

What it can read

Each of these answers for one connected tool, and the workspace is never an argument — it comes from the token, so no wording of a request reaches another customer's data.

It can also run four saved questions without you phrasing them: a weekly marketing operations review, a quarter review pack, what is broken right now, and whether sales is being sent good leads. And it can attach the overview of any connected tool as a document, without calling anything.

What it can propose, and what it can never do

With Also let it propose changes ticked, an assistant can put a change set in front of your team. That is the whole of the permission.

• It proposes. A person reviews the before-and-after list in MarqIQ and approves it, and a workspace admin applies it. This server carries no action that approves, applies or undoes anything.
• Two changes are catalogued: filling first-touch fields from what MarqIQ recorded when the person first arrived, and filling them from the person's most recent touch. Both only fill fields that are empty. Neither replaces a value.
• It works for HubSpot and Marketo. Any other tool is refused by name.
• The tool must already have first-touch fields mapped under Tool Operations › Field Mapping, or there is nowhere to write a first touch and the proposal is refused.
• One proposal per operation at a time. A second is refused while the first is waiting, so the same records are never proposed twice.
• The change set records the assistant that proposed it, the token's name and whose token it was.

A token without that permission is told so plainly: "This access token may read but not propose changes." Nothing else about it changes.

Check what a token has done, and revoke it

Every token in the workspace is listed under Tokens in this workspace, whoever made it.

1. 1

   Find the token in the list.

2. 2

   Click Revoke on its row.

3. 3

   Confirm the dialog.

   It reads "Any assistant using it stops working at once. This cannot be undone." Revocation takes effect on the next request; there is no cache to wait out.

When the assistant and a screen quote different figures

Both read the same analyses, so a difference is almost always the question rather than the answer. Check these three before you doubt either.

• The period. A screen you have filtered to last quarter and an assistant asked without a period are measuring different things. The assistant states the period it used in its reply.
• The connected tool. A workspace with two tools has two of every analysis, and an assistant asked in general picks one.
• When each was worked out. A screen left open shows what it read when it loaded. Reload it before comparing.

Where they still disagree, the screen is the one to raise a case about, and the assistant's reply is useful evidence: it names the tool, the period and the record count it rested on.