MCP Server for ChatGPT
Connect ChatGPT to your organization’s ConverSight data. Ask questions in plain English and get answers straight from live data — under your own sign-in, with the same permissions and row-level security that apply everywhere else in AI Insights.
Bring Your AI Assistant into ConverSight
The ConverSight app lets your team ask questions of the organization’s own data straight from ChatGPT, in plain English — no SQL, no dashboards, no exports. ChatGPT reads your datasets, runs the query for you, and explains what came back.
MCP, short for Model Context Protocol, is an open standard that lets AI assistants connect directly to outside systems and work with real, live data instead of only what they already happen to know. ConverSight speaks this standard, so ChatGPT can connect straight to your organization’s own data through the ConverSight MCP connector.
Each person connects under their own AI Insights sign-in, so ChatGPT sees exactly what that user is allowed to see. The same dataset permissions that apply everywhere else in AI Insights apply here too. Nothing is shared between people, and no data is copied out.
How it connects
ChatGPT connects to outside systems by adding a remote MCP server as a custom app (ChatGPT calls it an “MCP App” on the Plugins page). The AI Insights endpoint is a standard MCP server with OAuth sign-in, so no extra software is needed.
Before You Start
You need three things:
| Requirement | Description |
|---|---|
| ChatGPT on the web (chatgpt.com) | This guide was captured on a Plus account. On Business, Enterprise and Edu, a workspace admin may need to allow custom apps first — see Rolling Out to Your Team. |
| A ConverSight account | Belonging to the organization whose data you want to query. |
| Your MCP endpoint | Your own organization’s domain — see below. |
Your Organization’s MCP Endpoint
Click on the MCP Server module inside Data Studio to find your endpoint. The endpoint is your organization’s own domain with /mcp on the end. It is not a shared address — the domain is what tells the platform which organization you are signing in to, so it must be yours.
Setting Up the Connection
Setup runs in three stages, each covered in its own section below:
- Create the app: register ConverSight as a custom MCP app in ChatGPT, pointing it at your organization’s endpoint and the scopes it needs to query data.
- Sign in and allow access: authenticate with your own AI Insights credentials, then approve the permissions ChatGPT is requesting on the consent screen.
- Set tool permissions: choose how freely ChatGPT can use the connected tools, from asking before every action to approving low-risk reads automatically.
Creating the App
Click the Plugins menu item in the left-hand menu. Click the Add button in the top-right corner, then choose the Create MCP App option from the dropdown.
Choosing the Create MCP App option opens the popup directly. Fill in the required details as shown in the table below.
| Field | What to enter |
|---|---|
| Name | ConverSight |
| Description | Ask any questions of your ConverSight data (optional) |
| Connection | Stay on the Server URL tab and paste https://<your_org_domain>/mcp |
| Authentication | OAuth (the default) |
| Advanced OAuth settings → Scopes | query.read, metadata.read, metadata.sample_data |
| I understand and want to continue | Tick the checkbox, then click the Create button |
Open the Advanced OAuth settings section, then set the fields below.
| Field | What to do |
|---|---|
| Client ID, Client secret, Authorization URL, Token URL | Leave all four blank — ChatGPT discovers these from the server once you connect. |
| Scopes | Enter query.read, metadata.read, metadata.sample_data, one per line or comma-separated. |
Without the scopes above, ChatGPT only gets basic sign-in permissions: it can list your datasets, but every query fails with a “lacks query.read permission” message.
Tick the I understand and want to continue checkbox, then click the Create button.
Signing In and Allowing Access
Clicking the Create button brings up ChatGPT’s own connect popup for the new app, explaining that ChatGPT is strictly limited to the permissions you’ve set and that connectors can introduce risk if you don’t trust the server. Click the Continue to ConverSight button.
This sends you straight to your organization’s sign-in page, in the same browser tab. Signing in tells AI Insights who you are, so ChatGPT works with exactly the data your own account already permits.
Check the address bar shows your own organization’s domain, then enter your usual AI Insights credentials and click the Sign In button. The page says “Sign in to authorize ChatGPT”.
Click the Allow button on the consent page. It names ChatGPT as the app asking for access and lists what it is being granted; see Permissions for what each line covers. Those permissions set the outer limit of what ChatGPT may attempt — your existing dataset permissions still decide what data actually comes back.
You land back on the Plugins page. Click the Personal tab: ConverSight is listed under the Created by you section.
Tool Permissions
Click the app, then the more options icon beside the Try in chat button, and choose the Manage option. This page shows the account you connected with, the Permissions row, and, further down, the Tools section listing what ChatGPT loaded from AI Insights. Click the Permissions row to change how often ChatGPT asks before using a tool. See Tool Permissions for what to set.
Checking It Works
From the app page click the Try in chat button. You can also type @ConverSight in any new chat. The app appears as a blue tag at the start of your message.
Then ask the question List my curated datasets. You should get back a list of the datasets your account can reach, with ConverSight listed under Sources. From there, ask a real question — anything you would normally build a report for.
Tool Permissions
ChatGPT sets one permission policy for the whole app, on the Manage page’s Permissions row, not a policy per tool:
| Policy | What it means | Use it for |
|---|---|---|
| Always ask | ChatGPT asks before reading or making changes. | Tight control while your team is getting started. |
| Allow read-only tools | ChatGPT reads without asking, but asks before making changes. | Most teams — questions flow, changes are reviewed. |
| Allow low-risk tools (default) | ChatGPT automatically approves low-risk tools, but sensitive tools may be denied. | Fine to leave as is. |
| Allow all tools | ChatGPT uses tools without asking. Marked Elevated risk. | Not recommended. |
When ChatGPT does ask before a write action, it shows the exact request it is about to send. Read it before approving: a write action changes live AI Insights data.
Available Tools
You never call these by name. You ask a question, and ChatGPT picks the right tool. In our test connection, ChatGPT loaded the six tools below, grouped here by what each one does; the Manage page (see Tool Permissions) shows the list for your own connection.
Tools marked Write can change data or start work; everything else only reads.
Discovery Tools
Start here in any new conversation, to answer “what is available?” before asking anything about the numbers.
| Tool | R/W | What It Does | Ask It Like This |
|---|---|---|---|
list_datasets | Read | Lists the curated datasets your account can query, with the connector behind each one. | “List my curated datasets.” |
describe_dataset | Read | Describes one dataset — its streams, joins and pre-aggregations — and can return up to 10 sample rows. | “Describe supply_chain_v2 and show me a few sample rows.” |
list_streams | Read | Lists every table/stream in a connector’s catalogue. | “What fact tables are in the conn_csforesight connector?” |
Querying Tools
The everyday tool, for asking a question in plain English and getting a real answer back.
| Tool | R/W | What It Does | Ask It Like This |
|---|---|---|---|
query | Read | Runs a SQL query against a dataset and returns the rows, up to 1,000 in the preview. | “Total sales by each quarter.” |
Pipeline Tools
For work already defined as a pipeline in ConverSight, rather than a single query.
| Tool | R/W | What It Does | Ask It Like This |
|---|---|---|---|
validate_pipeline | Read | Checks a pipeline definition — SQL syntax, operators, wiring — without running it. | “Validate this pipeline before I save it.” |
run_pipeline | Write | Runs a multi-step pipeline definition end to end. | “Run the weekly replenishment pipeline.” |
If your ChatGPT tool list looks shorter than expected, check the Scopes under Creating the App, then remove the app and create it again.
Getting Better Answers
Six things that make a noticeable difference in practice.
1. Say “curated datasets” when you want the real ones.
- Asking for “datasets” also brings back connector entries, which are queryable but are not curated datasets. The narrower wording gets you a clean list.
2. Don’t ask for a connector’s whole stream list at once.
- Each entry carries its full schema, so a large lake can overflow what ChatGPT can read in one response. Ask for stream names, or one category at a time.
3. If ChatGPT doesn’t know the column names, ask it to just try a query.
- When a column name is wrong the engine replies with every valid column, so one deliberate attempt maps the whole table faster than asking for a description.
4. Refer to tables by their real names.
- “Join dimension_item to fact_sales_order_line_detail” works; generic placeholders do not.
5. Repeating a question is cheap.
- An identical question is served from cache and comes back much faster. Vary the wording if you want it re-run.
6. Ask for a chart.
- ChatGPT can plot what it queried — “show that as a bar chart by quarter”.
Start the message with @ConverSight so ChatGPT uses your data rather than web search or its own knowledge.
Troubleshooting
| If you see this | It means | Do this |
|---|---|---|
| There is no Create MCP App option in the Add menu. | Your workspace hasn’t enabled custom apps. | Ask your ChatGPT admin to allow it — see Rolling Out to Your Team. |
| Datasets list fine, but queries fail with “lacks query.read permission”. | The app was created without the Scopes. | Uninstall the app (see Disconnecting or Removing the App) and create it again with the Scopes from Creating the App. |
| The sign-in page shows a different organization. | The URL you entered was not your organization’s domain. | Uninstall the app and create it again using https://<your_org_domain>/mcp |
| Sign-in is refused even though the password is correct. | The account is not an active member of the organization that domain belongs to. | Sign in with an account that belongs to that organization. |
| ChatGPT answers without using your data. | It didn’t pick the app for that question. | Start the message with @ConverSight. |
| A query comes back as permission denied. | Your AI Insights account does not have access to that dataset. | Ask your AI Insights administrator to grant you read access to it. |
| You are asked to sign in again. | Saved access expired or was withdrawn. | Open the app’s Manage page and reconnect your account. |
| An answer looks out of date. | An identical question was served from cache. | Ask ChatGPT to run it again, or rephrase the question. |
Permissions
These are the items that can appear on the screen during Signing In and Allowing Access, and what each one covers.
| On the consent screen | What it allows |
|---|---|
| Verify your identity | Confirm that you are who you say you are. |
| View your profile information | Read your name and other basic profile details. |
| View your email address | Read the email address on your account. |
| Maintain access when you are not present | Keep the connection working between sessions, so you are not signed out after every conversation. |
query.read | List datasets, read their structure, and run read-only queries. |
query.write | Run pipelines, re-scan a connector, and upload a file as temporary data. |
metadata.read | Read catalogue information about connectors and streams. |
metadata.sample_data | Look at a handful of sample rows from a dataset. |
These permissions set the outer boundary of what the app may attempt. Within that boundary, your organization’s dataset permissions and row-level security still decide what data is returned — granting ChatGPT access does not widen anyone’s access to data.
How the Connection Registers Itself
When you click the Create button, ChatGPT registers itself automatically using the OAuth standard for dynamic client registration (RFC 7591) — which is why the Client ID and Client secret fields in the Advanced OAuth settings section are left blank and marked optional. Nothing has to be set up beforehand, and no client ID or secret is entered by hand.
A separate client identity is created for each connection and bound to the organization that the domain in your endpoint belongs to — which is why the endpoint must be your own organization’s domain. The approval you give on the consent screen can be withdrawn at any time.
Rolling Out to Your Team
On ChatGPT Business, Enterprise and Edu, owners and admins control who can use custom apps. Based on OpenAI’s documentation:
- Allow custom apps: Business — Settings › Apps › Advanced settings. Enterprise and Edu — grant it by role under Permissions & Roles › Connected Data.
- Create the app once for the workspace: Workspace settings › Apps › Create, with the same fields and Scopes as Creating the App.
- Choose which actions are allowed (Enterprise/Edu): use the Configure Actions button before publishing. Switching off the Write tools gives your team a read-only connector.
- Choose who gets it: use the Configure Access button to limit the app to specific roles.
Disconnecting or Removing the App
Open the Plugins › Personal tab, click the ConverSight app, then click the more options icon. The Manage option lets you disconnect or reconnect your AI Insights account and change permissions. The Uninstall option removes the app entirely; create it again from Creating the App if you need it back. The Delete plugin option goes further and removes the app definition itself, not just your connection to it.
Access also lapses on its own if unused for an extended period, which is why you may occasionally be asked to sign in again. That is expected.
Go to Plugins › Add › Create MCP App. Point it at https://<your_org_domain>/mcp with OAuth and the Scopes query.read, metadata.read, metadata.sample_data, click Create, then Continue to ConverSight, sign in and click Allow. Nothing needs to be set up beforehand.
Everyone sees only the data their own AI Insights account already permits. If something does not work, Troubleshooting covers the problems people actually hit.
Screens captured in ChatGPT on the web, September 2026. Admin controls are based on OpenAI’s help center articles “Developer mode and MCP apps in ChatGPT” and “Admin controls, security, and compliance for plugins and apps”. OpenAI updates these screens often.