Who Is Making This Request? Agent Auth 101 - Kamil Potrec, Keycard
MCP servers cannot distinguish agent requests from human CLI tools, creating an identity gap
“The MCP server doesn't really know if it's an agent, or, for example, Alice's command line interface, or any other tool.”
Keycard's product team outlines the layered credential problem in agentic systems: agents need user tokens, LLM API keys, and MCP server credentials simultaneously, yet MCP servers remain blind to whether a caller is an agent or a human tool. The industry is beginning to shift from static keys and forwarded PATs toward dynamic authentication, but MCP servers currently rely on service accounts that aggregate permissions as a workaround. This matters because the identity ambiguity at the MCP layer is a structural gap that complicates auditing, access control, and least-privilege enforcement in production agent deployments.