MCP Security

Model Context Protocol Elicitation and Consent Flow

MCP elicitation is a protocol primitive that lets a server request additional user input mid-session through the client, using a schema-defined request and an accept, decline, or cancel response. It is distinct from MCP's authorization process, and because the specification does not mandate audit logging, rate limiting, or enforcement, enterprises must build these runtime controls themselves before deploying elicitation-capable agents in production.

How the Elicitation Consent Flow Operates

At a protocol level, elicitation follows a fixed sequence between client and server, from initial capability declaration through to a validated, structured response.

  1. Capability Declaration

    The client declares elicitation support during initialization, establishing that it is able to surface mid-session requests to the user.

  2. Server Request

    The server sends a human-readable message together with a schema describing the structure of the input it needs.

  3. User Decision

    The user reviews the request and accepts, declines, or cancels it before any data is returned.

  4. Structured Response

    The client returns the user's decision, along with any validated data, back to the requesting server.

Runtime Controls to Implement for Elicitation Governance

The specification defines the elicitation message flow, but leaves enforcement to the surrounding deployment. The following controls are what enterprises need to add at runtime.

  • Require the client to display the requesting server, the requested schema, and explicit accept, decline, and cancel options before any data is returned.
  • Validate all user-submitted data against the declared schema before it leaves the client, rejecting malformed or out-of-schema submissions.
  • Apply rate limiting to elicitation requests from any single server to prevent excessive or repeated prompting during a session.
  • Log the request message, requested schema, user decision, and timestamp for every elicitation event, across all connected servers.
  • Define timeout and fallback behavior for cases where a client declares elicitation support but the user does not respond.
  • Restrict which users or roles are permitted to approve elicitation requests, particularly where responses might otherwise be handled automatically.

What Is MCP Elicitation?

Model Context Protocol elicitation is a primitive that lets an MCP server request additional information from a user through the client during an already active session, rather than only at the point of initial connection. Where a typical integration might collect user data once, up front, elicitation gives a server the ability to ask for structured input mid-task, when a new piece of information becomes necessary to complete a request. The elicitation request itself contains a human-readable message describing what is being asked and a schema that defines the expected structure of the response, allowing the client to validate what comes back before it is returned to the server.

Elicitation is transmitted using the same JSON-RPC based message format that MCP uses for other primitives such as tools and resources, so it fits into the protocol's existing request-response model rather than introducing a separate channel. It is worth distinguishing elicitation from MCP's authorization process. Authorization governs whether a client or agent is granted initial access to a server at all. Elicitation operates after that access has already been established, inside a session that is already running, and addresses a different problem: getting additional information from a human once an agent is mid-task.

Why Dynamic Elicitation Changes the Risk Model

Static permission models, where a user grants access once at connection time, create a single decision point that is comparatively easy to review, log, and govern. Elicitation introduces something structurally different: a runtime, event-driven consent checkpoint that can be triggered by a server at any point during an active agent session. Because the request originates from server logic rather than from a fixed schedule, an enterprise cannot predict in advance how many elicitation prompts a given task will generate or what data a server might attempt to request through them.

The specification directs servers not to use elicitation for sensitive information such as passwords or credentials. This is a specification directive rather than a technically enforced control. Nothing in the protocol itself prevents a server from sending an elicitation request that asks for sensitive data, and the client is the point where any resistance to that request would have to occur. Enterprises deploying MCP-connected agents should not assume this restriction is automatically respected by every server they connect to.

Multi-server and multi-agent deployments introduce a second complication: when several servers can each independently trigger elicitation, a user needs a consistent way to see which server is asking. Without clear attribution, it becomes difficult for a user to make an informed decision, and difficult for an audit team to reconstruct afterward which server initiated a particular prompt.

Bringing Elicitation Into Broader Runtime Governance

Elicitation sits inside a larger set of runtime decisions that enterprises need to govern once MCP-connected agents move into production: which tools an agent can call, what data it can access, and which servers it is permitted to reach in the first place. Because the MCP specification does not define a required audit logging format, retention policy, or enforcement mechanism for elicitation events, enterprises are left to build these controls themselves or rely on a governance layer positioned between agents and the MCP servers they connect to.

This is the layer where Trussed AI operates. Trussed AI provides runtime governance and security for enterprise AI agents, including audit logging, tool approval workflows, agent permissions, and least privilege controls for MCP-connected deployments. Rather than depending solely on server-side compliance with specification directives such as the restriction on requesting sensitive data through elicitation, a runtime governance layer can apply independent policy checks, log elicitation events consistently across every connected server, and give administrators visibility into which users approved which requests.

For governance leaders evaluating MCP-based agent deployments, the practical question is not whether elicitation is well specified. It is whether the surrounding runtime infrastructure captures, enforces, and audits it consistently before agents reach production.

Govern Elicitation and Consent Across Your MCP Deployments

See how runtime governance applies audit logging, permission controls, and policy enforcement to elicitation-driven consent events in MCP-connected agents.

Talk to an Expert