One of the trickiest parts of MCP in practice has been auth propagation. As soon as the agent backend invokes the MCP server instead of the client, the original user’s auth context disappears—tools that require the user's session_id (or equivalent) suddenly only see a generic token. We ended up needing a pattern for:
- M2M-issued short-lived tokens for backend → MCP calls
- Per-request user metadata injection so tool calls can still act on behalf of the user
- Consistent OAuth2 / Okta validation so both layers trust each other
- M2M-issued short-lived tokens for backend → MCP calls
- Per-request user metadata injection so tool calls can still act on behalf of the user
- Consistent OAuth2 / Okta validation so both layers trust each other
Was looking for this standarization.