MCP Server for Claude
The ConverSight MCP connector empowers organization administrators to connect any MCP-compatible client directly to their organization's ConverSight data in about five minutes per person, letting every team member ask questions in plain English and get answers straight from live data — no SQL, no dashboards, no exports.
This documentation uses Claude as its worked example throughout — the screens, steps, and screenshots shown are Claude Desktop's. The connector itself works the same way with any client that supports MCP; only the exact screens differ.
Bring Your Own AI Assistant into ConverSight
Connecting the AI assistant a team already uses to ConverSight lets it work with that organization's real data, right inside the conversation already underway. A question asked there returns an answer pulled from live datasets — not a guess, not a stale export, and not a spreadsheet to track down first.
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 any MCP-compatible client can connect straight to an organization's own data through the ConverSight MCP connector.
Before You Start
Administrators need three things before adding the connector:
| Requirement | Description |
|---|---|
| An MCP client | Installed and signed in on the machine that will be used. This guide uses Claude Desktop as the example client. |
| A ConverSight account | Belongs to the organization whose data the connector should reach. |
| An MCP endpoint | Points at the organization's own ConverSight domain — see below. |
In Claude Desktop, if the account is an organization account, only an admin of that organization can add the MCP connector. Members without admin rights will not see the option to add a custom connector. Admin controls like this vary by MCP client.
Your Organization's MCP Endpoint
Click on the MCP Server module inside Data Studio to find the endpoint. The endpoint is the organization's own ConverSight domain with /mcp on the end. It is not a shared ConverSight address — the domain is what tells ConverSight which organization a person is signing in to, so it must belong to that organization.
- Ask in Plain English: No SQL, no dashboards, no exports. Your AI assistant reads your curated datasets, runs the query, and explains the result in plain language.
- Respects Your Permissions: Each person connects under their own ConverSight sign-in. Your AI assistant sees exactly what that person is allowed to see, with the same row-level security that applies everywhere else in ConverSight.
- Six Built-In Capabilities: Explore your datasets, run and re-run analyses, and confirm the results are accurate, all as soon as you connect, with nothing to set up first.
- Your Data Stays in ConverSight: Your data stays securely within ConverSight, kept private to you. You're always in control, free to adjust the access you granted on the consent screen at any time.
How It Works
Connecting takes four steps, covered in full detail in the sections below:
- Get the endpoint: Open MCP Server in ConverSight and copy the organization's connection address.
- Add it to the AI assistant: Paste the endpoint into the client's connector settings and give it a name the team will recognize.
- Sign in and review access: Sign in with a ConverSight account and check the permissions being requested before allowing the connection.
- Set how much it can do unprompted: Choose which actions it can take on its own and which should check first — see Tool Permissions.
Once connected, just ask — the AI assistant finds the right dataset, retrieves the answer, and brings it straight back into the conversation.
Accessing the MCP Server
ConverSight MCP is added from inside an MCP client as a custom connector, because it points at an organization's own domain rather than one of the client's pre-built entries. The steps and screenshots below show Claude Desktop as the example client; other MCP clients follow the same general pattern from their own connector or MCP server settings.
Step 1: Click on the account name at the bottom of the left-hand sidebar to open the account menu, then select Settings.
Step 2: Click on Connectors in the left-hand menu of Claude Desktop's Settings page to open the list of connectors that copy of Claude Desktop can use.
Step 3: Click on Add at the top right of the Connectors page, then select Add custom connector to open the connector popup.
Setting Up the Connection
Connecting an MCP client to ConverSight takes two stages, shown here in Claude Desktop. First, add the connector by entering its endpoint and confirming how it authenticates. Then sign in and allow access, granting the client the permissions it needs. Everything happens inside the client apart from the sign-in, which opens in the browser. Once connected, tool permissions decide how freely the client may use each capability.
Adding the Connector
Creating the connector takes two steps in Claude Desktop: entering its name and endpoint, then confirming its authentication settings.
Step 1: Enter the name and the endpoint
Type a name the team will recognize, such as ConverSight, then paste the endpoint into Remote MCP server URL and click on Add to create the connector. Leave the OAuth fields under Advanced settings empty; they fill in automatically.
Clicking Add either creates the connector immediately or opens a second step to confirm authentication settings, depending on the Claude Desktop version. If the connector is created immediately, continue to signing in and allowing access.
Step 2: Confirm the authentication settings
Clicking Add moves the popup to a second step, showing the sign-in requirements Claude detected for the server. Leave Required when the server asks and No client ID — register one automatically selected, since these are the settings that work for most servers, then click on Add again to finish creating the connector.
Signing In and Allowing Access
The connector is now ready to connect. Signing in tells ConverSight who the person is, giving the client access to exactly the data that account already has.
Step 1: Start the sign-in
Click on Connect button in the connector's page to open the organization's ConverSight sign-in page in the browser.
Step 2: Confirm you want to continue
The browser shows a one-time confirmation page before it hands off to ConverSight. Click on the Continue Connecting button to proceed to the sign-in page.
First-time users are prompted to sign in to their Claude account first. The browser then continues automatically to the organization's ConverSight sign-in page.
Step 3: Sign in to ConverSight
After being directed to the login page, check that the address bar shows the organization's own domain before typing anything. Once checked then enter the usual ConverSight credentials and then click on the Sign In button.
Step 4: Review and allow access
Review the permissions listed on the consent page, then click on the Allow button to grant the client those permissions. The account's own dataset permissions still decide what data actually comes back, so allowing access here never grants more than the account already has.
See Permissions for what each one covers.
Step 5: Return to Claude Desktop
Click on Open Claude in the browser popup to hand control back to the desktop app, then close the browser tab.
Tool Permissions
Tool permissions let administrators choose exactly how freely the client can act, from asking before every use to running on its own. Every tool starts on a needs-approval policy, so nothing runs unprompted until an administrator changes it; Claude Desktop's three-tier policy is used as the example below.
Click on the approval dropdown at the top right of the tool list to set how the client may use each tool. See Tool Permissions Reference for what to set on each one, and Available Tools for what each tool actually does.
Tool Permissions Reference
Each tool can be set to one of three policies. Claude Desktop names them as follows; other MCP clients expose the same three concepts, sometimes under different labels:
| Policy | What it means | Use it for |
|---|---|---|
| Needs approval | The client asks every time before using the tool. | Anything that writes or changes data. |
| Always allow | The client uses the tool without asking. | Read-only tools used constantly. |
| Never allow | The tool is hidden from the client entirely. | Capabilities the team should not use. |
A good starting point: set the read-only tools to Always allow so day-to-day questions flow without interruption. Available tools marks which tools those are.
Leave every Write-marked tool on a needs-approval policy, whatever client is being configured. Those tools can change data or start work, so a person should see the request before it runs.
Checking It Works
Start a new conversation in Claude Desktop and ask "List my curated datasets". Claude asks permission to use the tool the first time. Choose Always allow for read-only tools like this one — Claude will not ask again.
The response lists the datasets that account can reach. From there, ask a real question — anything that would normally be built into a report.
Asking "List my datasets" returns the same style of answer as "List my curated datasets," filtered to the curated set — see Getting better answers.
Available Tools
The ConverSight MCP connector gives the AI assistant everything it needs to explore, query, and act on an organization's data, grouped below by what each one does. A person never calls a tool by name — asking a question in plain English is enough for the AI assistant to pick the right one — but administrators can use this reference to see exactly what the connector is capable of, and the sample questions show the phrasing that reaches each tool.
Tools marked Write can change data or trigger work; everything else only reads. The short names below are the ones ConverSight uses in its own logs, so a name in a log traces back to a row in this reference. MCP clients typically surface the same tools in plain words in their own tool-permissions panel, such as Describe dataset or List datasets; Claude Desktop's panel is shown in the screenshots above.
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 datasets the account can reach, and counts how many are curated datasets rather than whole-connector entries. | "List my curated datasets." |
describe_dataset | Read | Describes one dataset — its streams, joins and pre-aggregations, and when its data was last synced — and can return up to 10 sample rows. | "Describe supply_chain_v2 and show me a few sample rows." |
list_streams | Read | Lists the tables and views one dataset publishes — the curated surface the team chose to expose, not the raw storage behind it. | "What tables are in supply_chain_v2?" |
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 |
|---|---|---|---|
run_pipeline | Write | Runs a multi-step pipeline definition end to end. | "Run the weekly replenishment pipeline." |
validate_pipeline | Read | Checks a pipeline definition — SQL syntax, operators, wiring — without running it. | "Validate this pipeline before I save it." |
Earlier releases exposed a wider set of tools, including background jobs, file upload, scenario branches, and writeback. Those capabilities still exist in ConverSight; they are no longer exposed over MCP, so a client reports that it has no such tool.
Getting Better Answers
This guide outlines quick habits that get faster, more accurate answers from the AI assistant. Think of it like working with a human analyst — the clearer the question, the better the answer.
1. Say "Curated Datasets" When the Real Ones Are Wanted
Asking for "datasets" also brings back connector entries, which are queryable but are not curated datasets. The narrower wording gets a clean list.
2. Aggregate in the Question
"Total revenue by region last quarter" comes back complete. "Show me last quarter's order lines" stops at the 1,000-row preview and can still look finished. Let the engine do the counting instead of asking for individual rows.
3. Let the Engine Guess the Column Names
Dataset descriptions are often thin. If the AI assistant 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. Repeat a Question When the Answer Looks Stale
An identical question is served from cache and comes back much faster. Vary the wording to force a genuine re-run.
6. Ask for a Chart
The AI assistant can plot what it queried — "show that as a bar chart by quarter."
Troubleshooting
| If you see this | It means | Do this |
|---|---|---|
| The connector is listed but says it is not connected. | The sign-in step was never completed. | Open the connector and click Connect (see Signing in and allowing access). |
| The sign-in page shows a different organization. | The URL entered was not the organization's domain. | Remove the connector and add 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. |
| The client reports that no tools were found. | The connection completed but the tool list did not load. | Disconnect and reconnect. If it persists, contact ConverSight support. See Available Tools for what should be there. |
| Every single action asks for approval. | New connectors default to Needs approval for all tools. | Set the read-only tools to Always allow in Tool permissions reference. |
| A query comes back as permission denied. | The ConverSight account does not have access to that dataset. | Ask a ConverSight administrator to grant read access to it. |
| The user is asked to sign in again after about a week. | Normal — the connector's saved access expires. | Open the connector and click Connect again. |
| An answer looks out of date. | An identical question was served from cache. | Ask the client to run it again, or rephrase the question. |
Permissions
These are the items listed on the consent page during signing in and allowing access, and what each one covers.
| On the consent screen | What it allows |
|---|---|
| Verify your identity | Confirm that the signed-in person is who they say they are. |
| View your profile information | Read the person's name and other basic profile details. |
| View your email address | Read the email address on the account. |
| Maintain access when you are not present | Keep the connection working between sessions, so the person is not signed out after every conversation. |
| Run queries and view your dataset data | Run queries against datasets and read back the results. This line covers dataset querying in general — confirm with a ConverSight administrator whether it also extends to write actions such as running pipelines. |
| View dataset metadata | Read catalogue information about connectors and streams. |
| View sample rows from your datasets | Look at a handful of sample rows from a dataset. |
Earlier versions of this consent screen listed query.read and query.write as two separate lines. The current screen combines dataset querying into the single line above — confirm with ConverSight if a rollout needs read and write access distinguished.
These permissions set the outer boundary of what the connector may attempt. Within that boundary, the organization's dataset permissions and row-level security still decide what data is actually returned — granting the connector access does not widen anyone's access to data.
How the Connection Registers Itself
Nothing needs to be registered with ConverSight in advance. Clicking Connect during signing in and allowing access makes the client register itself with ConverSight automatically, using the OAuth standard for dynamic client registration (RFC 7591).
A separate client identity is created for each installation, and it is bound to the organization that the domain in the endpoint belongs to — which is why the endpoint must be the organization's own domain. This is also why the OAuth Client ID and Secret fields are left empty when entering the name and endpoint: they are issued during registration rather than entered by hand.
Because the connector registers itself rather than being pre-approved, ConverSight asks for approval of its access on the consent page, and that approval can be withdrawn at any time.
Disconnecting or Reconnecting
In Claude Desktop, open Settings › Connectors, select the ConverSight connector, and use Disconnect. Other MCP clients expose the same action from wherever they manage connectors or MCP servers. This ends the client's access immediately; the connector stays in the list so it can be reconnected with the Connect button. To remove it entirely, use the menu beside Disconnect.
Access also lapses on its own if unused for an extended period, which is why a person may occasionally be asked to sign in again. That is expected and needs no action beyond clicking Connect.
Rolling this out does not widen anyone's access — every person still sees only the data their own ConverSight account already permits.