Connect a system to the agent

Collections handle stable knowledge, such as policies, manuals and catalogs. They don't handle what changes by the minute: an order's status, a stock balance, the ticket opened yesterday.

For that the Agent needs a tool: something it calls at the moment of the question. This guide decides which path to take and sets up the connection.

Collection or tool

The question is whether the data has a shelf life.

• Collection: the content is valid for weeks or months. Indexing is cheaper and faster at answer time.
• Tool: the data has to be current. Each question triggers a real query to the system.

The same Agent can have both: the returns policy in the Collection, the order status through the tool.

Integration or MCP

Once you've chosen the tool path, there are two ways:

• Integration, when the service is in the catalog the platform already knows. Less configuration, maintained by the platform.
• MCP, when the system is internal or specific to your business, or when you need to control exactly which tools are exposed.

The path, end to end

Both follow the same three-screen sequence, handled by different roles:

1. In the organization, someone with the Superadmin or Admin role registers the MCP server or the integration in the catalog, with the access credential.
2. In workspace management, the item is associated with the workspace that will use it.
3. In the workspace, it is installed, with its own name and configuration.

Hand it to the agent

Installing in the workspace still doesn't hand anything over. In the Agent form, check the item in the MCPs or Integrations section.

In the MCPs section, also adjust the maximum iterations: how many times the Agent can call tools in one answer. The default suits most cases; raise it only if the Agent is wrapping up before finishing the job.

Write the prompt with the tool in mind

Standing instructions need to say when to use the tool. Without that, the model tends to answer from memory instead of querying.

For any question about the status, deadline or amount of an order,
query the system before answering. Never estimate this data.

When the tool isn't called

In order of frequency:

• It isn't checked in the Agent form. Installing in the workspace isn't handing it over.
• The prompt doesn't tell it to use the tool, and the model answered without querying.
• The iteration limit was reached before the calls were completed.
• The server is down. The failure shows up in the traces, under Observability.

Traces are how you tell the first two cases from the last two: they show whether the call actually happened.

Security

An MCP gives the Agent the ability to act on real systems, not just read them. Before releasing a server to teams, check the result of the automatic analysis in Security scanner.

Next steps

• Register the server in MCPs.
• Or use a ready-made connection in Integrations.