Connect an assistant

Point your own AI assistant at your Bunk3r data and ask it questions directly. It takes a couple of minutes, it can only read, and you can disconnect it at any time.

Who can do this: owners and admins. Connecting an assistant is an administrative act, like inviting a teammate. If you are a closer and want to ask about your own book, ask an owner to create a token for you.

Claude (claude.ai)

Custom connectors are a paid-plan feature on claude.ai. If you do not see the option, that is why.

In Claude, open Settings → Connectors → Add custom connector.
Paste this URL:
```
https://app.bunk3r.ai/api/mcp
```
Leave the rest alone. Claude detects how to sign in and registers itself — the authentication and OAuth options it offers should both say Detected, and the request-headers box stays empty.
Finish, and Claude sends you to Bunk3r to approve it. Sign in if asked.
On the approval screen, tick the organizations this assistant may read, and press Allow. If you run several floors, tick the ones you want — see several organizations below.

That is the whole setup. Ask it something from the list below to confirm it works.

ChatGPT

ChatGPT reaches connectors like this one through developer mode, which is off by default and has to be turned on before the option to add one appears. It is a beta feature and availability depends on your plan.

In ChatGPT, open Settings → Connectors → Advanced settings and turn on Developer mode.
Back in Connectors, create a new connector and paste the same URL:
```
https://app.bunk3r.ai/api/mcp
```
Choose OAuth for authentication. ChatGPT registers itself, then sends you to Bunk3r to sign in and approve — the same screen Claude uses.
Tick the organizations it may read, press Allow, and start a chat with the connector enabled.

ChatGPT will warn you that connectors can take actions on your behalf. That warning is about connectors in general; this one has no tool that writes, so there is nothing for it to do beyond reading the figures you approved. Developer mode is also where ChatGPT allows write-capable connectors, which is why the warning is worded the way it is.

Claude Code, Cursor, Codex and others

Any client that speaks MCP connects the same way. Clients that run on your own machine handle sign-in in a browser window and keep the credential locally.

claude mcp add --transport http bunk3r https://app.bunk3r.ai/api/mcp

The first call opens a browser to the same approval screen. For a client that cannot sign in this way — a script, or a tool with no browser — create an API token instead: in the dashboard, People Access → Assistant access → Create token. The token is shown once, and it carries the access of whoever created it.

Several organizations

One connection covers every organization you tick when you approve it, so an agency connects once rather than once per floor.

When a connection covers more than one, every question has to say which one it means — and an assistant that does not name one is told to pick rather than allowed to guess. In practice you just say it: “how did Blueprint do yesterday?” Ask “compare my floors this month” and it uses the roll-up instead, which counts each organization in its own timezone and refuses to add unlike currencies.

Adding a client later? The organizations are fixed when you approve them, so a new one is not readable until you reconnect and tick it. That is deliberate: nobody should gain access to a company's revenue without somebody approving it.

What to ask first

These confirm the connection works and show what it is good at:

What organization am I connected to, and what data do you have?
What's our revenue for the last 30 days?
Why is cash collected today what it is? Break it down.
Who are my top closers this month, and how much has each closed?
How many calls are booked in the future?
Which leads from last week never got a call?
What are we owed in commissions, and why doesn't it match collected revenue?

The one worth knowing about is “why is this figure what it is?” An assistant can pull the individual payments, deals and calls behind any total, which is the thing a dashboard cannot do for you.

What it can and cannot do

Reads only. Nothing it can call changes a deal, a target, a commission rate or a connected account.
Sees what you see. It inherits the access of the person who connected it, checked fresh on every request. Change somebody's role and their assistant narrows with them; remove them from the organization and it stops working immediately.
Stays in the organizations you ticked. It cannot reach another company's data, including one you belong to but did not approve.
Sends data to your assistant's provider. Whatever it reads is processed by whoever runs the assistant, under their terms. That is worth knowing before connecting one to revenue data.

Disconnecting

In the dashboard: People Access → Assistant access. Every connected assistant is listed with who approved it, which organizations it covers and how often it has been used. Press Disconnect and it stops reading within seconds.

Disconnecting cuts the assistant off from every organization that connection covers, not just the one you are looking at. The list says how many, so there is no surprise.

Next: what an assistant can actually read, or what the figures mean.

Something here wrong or missing? Tell us — these pages are part of the product.