The team exposed the search through an MCP server with an HTTP /mcp endpoint. An agent calls indexer_search through JSON-RPC with a query and filters; the service returns document fragments and links.
The article separates permission to use the search tool from permission to read a document. For the Confluence integration, Indexer retrieves candidates and passes page identifiers plus the user identity to an ACL gateway, which checks Confluence space permissions and page restrictions before text reaches the model. The team says its initial candidate set is three times the requested result count, with up to three retrieval rounds and a cap of 100 candidates per request, to replace results removed by access checks.
For denied documents, the implementation retains a title and link in denied_results while excluding content. Runity notes that whether even the existence of a restricted document may be disclosed requires a company-specific decision. The source describes a case study and a synthetic test of access restrictions; it does not establish comparative retrieval quality, production-scale performance, or an independent security assessment.