Use a Data Collection from another AI tool
Every Data Collection publishes an MCP endpoint. Point Claude, Cowork, or any other MCP client at it, and that tool can ask your Data Collection questions and run your team's verified queries, without the data leaving your warehouse.
Audience: builders and business users
What this gives you
People do not do all their thinking in one tool. The MCP endpoint means the work you did building a Data Collection, the sources, the vocabulary, the queries your team approved, is available wherever someone is already working.
The connected tool gets the same guardrails you built. It asks questions through the Data Collection, so answers come back in your business vocabulary rather than raw column names. It runs verified queries, so the SQL is the SQL your team approved. Queries execute against the source, so nothing is copied anywhere new.
Data Collections only. The MCP endpoint is published for Data Collections. Ordinary Projects do not publish one: they expose a project-scoped REST API instead, which you will find in the same settings page.
Connect a client
Open the Data Collection's settings
Go to the Data Collection and open its settings, then find the MCP endpoint card.
Copy the endpoint URL
Use the copy control on the URL field. The URL is specific to this Data Collection: each one has its own.
Add it to your client
In Claude, Cowork, the MCP Inspector, or whatever client you are using, add a custom connector and paste the URL.
Authenticate
Clients that support MCP OAuth discover the handshake from the URL itself: you approve the connection in the browser and no key changes hands. For clients that do not, create a key in the REST API section of the same settings page and send it as an Authorization: Bearer header.
Check it connected
The MCP endpoint card shows a clients connected badge and lists each connected client below it.
What the client can do
The endpoint exposes four tools, plus one typed tool per published verified query:
| Tool | What it does |
|---|---|
ask_question | Ask the Data Collection a question in plain language. It resolves the question against your vocabulary and runs the query. |
list_verified_queries | List the verified queries this Data Collection publishes. |
run_verified_query | Run one of them by name, with arguments. |
get_intent_run | Read back a run that was started earlier. |
verified_<slug> | One typed tool per published verified query, so the client can call it directly with named arguments. |
The typed tools are what make a connected client good at this. Rather than asking the model to compose a query, it picks the approved query by name and fills in the arguments, and the client shows it as "Verified query: " followed by the title your team gave it.
What you should know before you connect one
Only approved verified queries are exposed. A query in draft is not visible over the endpoint. Publishing a verified query is what puts it in front of connected tools, so treat publishing as the decision it is.
Connections are per client, and revocable. Each connected client is listed by name. Revoking one ends its refresh tokens immediately and it has to reconnect.
Answers leave the platform. The client asked the question, so the answer goes back to the client. Whatever your organization's policy is for that tool applies to the data it receives.
An API key is a standing grant. A key sent as a bearer token works until it is revoked, and it is not tied to a person. Prefer OAuth where the client supports it, and keep keys to clients that cannot do the handshake.
Technical details
| Transport | Streamable HTTP, MCP revision 2025-06-18 |
| URL | {your deployment}/api/v1/project-api/projects/{project_id}/mcp |
| Authentication | MCP OAuth, or an API key as Authorization: Bearer |
The endpoint is the same one your administrator sees, so if a client cannot reach it, the thing to check first is whether your deployment's network endpoints allow the client in.
This is the outbound direction. Connecting other people's MCP servers into the platform, so your Projects can use their tools, is a different thing. See MCP servers for that.