Skip to main content

How can we secure Salesforce hosted MCP server that it has less record access provided to Claude users - based on record fields such as country or product line, similar to attribute based access control - than if they were logging into Salesforce from a browser? 

 

We do not want to change the record sharing and access rules in Salesforce for logged in users but want to restrict it for Claude or other AI tool users.  

 

What options would you recommend? 

 

 

10 respostas
  1. 9 de set., 14:17

    Hi @Istvan Szegedi yes, you can accomplish this in DataGrout in a few different ways:  

      

    1.    In DataGrout’s Salesforce MCP Server, there are a series of data retrieval tools. With over 700 Salesforce tools, those get quite granular. For e.g., in the case of the Opportunity object, your agent can retrieve relevant Opportunity details via the “get-opportunity*” tools.   

      

    2.    In case, your agent does not find a tool that matches your retrieval criteria (say, by region/country/custom-field), you can use DataGrout’s ToolSmith to add relevant (custom) tools to your MCP Server and have DataGrout expose that MCP Server/tool to your agent (while withholding the generic SOQL tool from the MCP Server so the agent does not compose its own queries to bypass the tool restrictions).  

      

    3.    This will work for aggregations/summaries as well via the corresponding filter being added to the SOQL query within a (region-specific) MCP tool such as “SELECT SUM(amount) FROM opportunity WHERE country = ‘United States’”. The agent could simply ask for “SELECT SUM(amount) FROM opportunity” but what DataGrout will execute (via the forementioned MCP tool) is “SELECT SUM(amount) FROM opportunity WHERE country = ‘United States’”.  

      

    4.    We also have policy guard-rails in the DataGrout MCP Gateway that can be used to restrict agent access to certain record sets or operations independent of the Salesforce user’s privileges via monotonic policy cascades.   

      

    Last but not least, I’m having one of the DataGrout product specialists reach out to you to schedule a call and see how we might be able to further help with this or additional use cases.

0/9000