MCP connects the tools.
What has to be decided around it?

The role of the protocol

The Model Context Protocol provides a standardized way to connect AI applications to data sources and tools. It makes a common interface easier; it does not replace the business application, its access rights or the quality of the information.

A useful comparison: the protocol organizes the connection. It does not decide, on behalf of your company, that an assistant may change an order or look at all customer records.

Reading and writing are two different needs

Fictional example: checking the status of an inventory item to prepare a draft reply. The tool exposes one precise read operation: a reference, a warehouse and a dated result. For this use, there is no need to also give it the ability to change quantities or delete products.

If the use moves toward an action, describe the operation expected, any validation step, the trail to keep and the behavior on failure. A “success” response from the connector must correspond to a state that can be checked in the target tool.

Checks before connecting

1. Identify where the MCP server comes from and the functions it actually exposes.
2. Check the account used and restrict its rights where needed.
3. Examine which data goes to which service.
4. Test a valid call, a missing piece of information and a refused access.
5. Check that access can be revoked and how things behave if the connector is unavailable.

Instructions found in a page or a document must not automatically become orders for the agent. The data consulted and the execution permissions are two levels to handle separately.

A protocol does not mean adopting a whole vendor

The same need can sometimes be covered by an existing API or a simpler integration. The choice depends on the client applications, the environment and the maintenance work. Saphan can compare these routes on a defined business workflow, rather than multiplying connectors with no use.