Skip to main content
AI Security

MCP Elicitation: Keep Sensitive Steps Outside the Agent

How URL-mode elicitation creates safer boundaries for payments, credentials, consent, and third-party authorization in agent workflows.

Baddula Pranathi
Baddula Pranathi · 6 min read
MCP Elicitation: Keep Sensitive Steps Outside the Agent

Some steps should never pass through an agent conversation: entering a password, approving a payment, completing a third-party OAuth flow, or sharing a secret. URL-mode elicitation gives the server a way to direct the person to a trusted out-of-band page and later resume the workflow from a result.

Separate assistance from authority

The agent can explain why a step is needed and what will happen next. The sensitive page should own authentication, consent, validation, and submission. The agent receives only the minimum status required to continue.

  • Show the destination and purpose before the user opens it
  • Use a unique, short-lived correlation identifier
  • Do not place credentials, payment details, or authorization codes in chat
  • Bind the callback to the initiating user, session, and exact request
  • Expire abandoned requests and make retry behavior explicit

Design the return path

A secure handoff still needs good product design. Tell the user when they can close the external page, show whether the action succeeded, and let the workflow recover cleanly from cancellation or expiry.

Use elicitation sparingly

Every handoff interrupts the flow. Reserve it for moments where a direct user interaction materially improves security, consent, or compliance, and keep ordinary low-risk inputs inside the main experience.

Primary sources

First-party documentation and announcements used to ground this field note.

MCPElicitationPaymentsOAuthSensitive Data
Baddula Pranathi
Baddula PranathiFull-Stack Developer · Atyuttama