Home / Docs / MCP

MCP behavior

Why this model fits MCP clients better than flat env exposure.

KeyLore works well with MCP-style interactions because MCP clients often need a structured way to discover resources and invoke tools. In this model, credentials are not dumped into process environment state for the agent to read.

  • Expose enough metadata for the client to distinguish credentials by service, purpose, and environment.
  • Do not rely on raw secret names alone; make the intended use explicit in the metadata.
  • Keep the AI-visible layer descriptive but minimal so selection is accurate without oversharing.
  • Validate the workflow by confirming the client can complete the task without ever seeing the underlying token value.

Model-visible contract

What the agent sees.

MCP clients are especially sensitive to interface design. If the only interface to credentials is a flat list of env vars, the agent must infer intent from weak signals. KeyLore gives that interaction a better structure: searchable metadata for discovery, brokered access for use, and a clearer separation between model context and secret material.

Visible over MCP

{
  "service": "github",
  "allowedDomains": ["api.github.com"],
  "permittedOperations": ["http.get"],
  "userContext": "Human explanation",
  "llmContext": "Use for repo metadata and rate-limit reads"
}

Not returned

{
  "rawToken": "not exposed",
  "secretBinding": "not exposed",
  "envDump": "not part of the flow"
}