Connect data and knowledge through MCP
MCP lets Luxiva use capabilities exposed by an external service. Connect first, define a logical dataset with a clear purpose, then select it for an agent. You do not have to create a metric for every number. Compatibility depends on the provider's protocol, authentication and actual read-only capabilities; not every MCP service or format has been verified.
Before you start
Workspace owners and collaborators can manage connections and prepare and save datasets. A collaborator is not a platform administrator. Members and shared recipients use authorized agents rather than manage onboarding. Ordinary management permission does not grant owner-only background automation authority.
In Data → Add → Connection → MCP, enter the provider's address and supported authentication details. Never put credentials in chat, a business purpose or shared knowledge. A working connection or capability listing does not prove that the intended data is readable.
Default path: data onboarding Copilot
- Add a dataset on the MCP connection and use the default Copilot entry. Create separate datasets for different repositories, projects or semantic models rather than replace an existing working scope.
- Describe your purpose, such as “review course registrations this school term,” or choose exploration. Include the objects, period, definitions and restrictions; do not assume the system knows your business identifiers.
- Review the specific read-only capabilities and explicitly confirm read access. Preparing a plan is not a data read; metadata discovery is not query evidence.
- Start bounded exploration. The interface shows budgets and progress. Answer follow-up questions about missing projects, models, periods or other necessary inputs. Cancellation or reaching a limit is not a successful validation.
- Review the scope, proposed read capabilities, verified findings and unknowns. Listing objects, reading definitions or seeing samples does not establish a complete population or support for every filter and period.
- Acknowledge the read-only scope and click Save. Wait for the saved state. A draft or ready preview is not a completed save.
- Inspect the dataset in Data Mart, then explicitly select it when creating or editing an agent. A connection does not authorize itself, and sibling datasets on that connection are not included.
Adding a dataset configures an independent logical scope for on-demand reads, not a bulk copy of the service. Reads transfer necessary data, and the platform may retain definitions, validation evidence, profiles and run outputs. This is not a promise of no transfer or no retention.
Advanced path: manual preparation and checks
Switch to the manual wizard when you need detailed control. Set a purpose, prepare relevant checks, review capabilities and inputs, explicitly authorize the read, then run, review and save. Preparation is not execution. A successful directory or definition check cannot substitute for validating the intended query. Do not remove filters, switch metrics or change periods merely to make a check pass.
Troubleshooting
- Permission: ask an owner or collaborator to manage setup; ask the Builder to check the agent's exact source selection.
- Authentication: renew credentials in connection settings and recheck the intended scope. Do not send credentials to Copilot.
- Missing input: supply the requested business parameter rather than guess identifiers or enum values.
- Provider refusal: check the original goal and accepted inputs. A provider can reject a query even when connectivity works.
- Failed, cancelled or unfinished: review the safe explanation and available evidence, correct the cause and explicitly retry. Partial evidence is not a pass.
- Changed source or revoked access: old evidence may be stale; recheck the current definition. The dedicated Zen integration is retired. A compatible provider can connect through standard MCP without restoring that old integration.
Continue with Data Mart and shared profiles, Create an agent and Knowledge review.