Elicitation - AgentFront

Elicitation allows tools and agents to request interactive user input during execution. This enables approval workflows, multi-step confirmations, OAuth flows, and any scenario where a tool needs user input before continuing.

This feature implements the MCP Elicitation specification. FrontMCP handles all protocol details automatically.

Why Elicitation?

In the Model Context Protocol, elicitation fills a critical gap:

Aspect	Tool	Elicitation
Direction	Model triggers, server executes	Server requests, user responds
Interaction	One-way (model → server)	Two-way (server ↔ user)
Use case	Actions, calculations	Confirmations, approvals, OAuth
Blocking	No	Yes (waits for user response)

Elicitation is ideal for:

Approval workflows — confirm destructive operations before execution
Multi-step forms — collect additional information during tool execution
OAuth flows — redirect users to external authentication (URL mode)
Human-in-the-loop — require explicit user consent for sensitive actions
Disambiguation — ask users to clarify ambiguous requests

Basic Usage

Class Style (ToolContext)

Use this.elicit() within any tool that extends ToolContext:

import { Tool, ToolContext } from '@frontmcp/sdk';
import { z } from '@frontmcp/sdk';

@Tool({
  name: 'delete-account',
  description: 'Permanently delete a user account',
  inputSchema: {
    userId: z.string(),
  },
})
class DeleteAccountTool extends ToolContext {
  async execute({ userId }: { userId: string }) {
    // Request user confirmation before proceeding
    const result = await this.elicit(
      `Are you sure you want to permanently delete account ${userId}? This cannot be undone.`,
      z.object({
        confirmed: z.boolean().describe('Confirm deletion'),
        reason: z.string().optional().describe('Reason for deletion'),
      })
    );

    // Handle user response
    if (result.status !== 'accept') {
      return { cancelled: true, reason: `User ${result.status}d the request` };
    }

    if (!result.content?.confirmed) {
      return { cancelled: true, reason: 'User did not confirm deletion' };
    }

    // Proceed with deletion
    await this.deleteUser(userId, result.content.reason);
    return { deleted: true, userId };
  }
}

Class Style (AgentContext)

Agents are LLM-powered autonomous units that can also use elicitation. Override execute() to add custom logic like approval workflows:

import { Agent, AgentContext } from '@frontmcp/sdk';
import { z } from '@frontmcp/sdk';

@Agent({
  name: 'expense-processor',
  description: 'Processes expense reports with AI categorization and manager approval',
  systemInstructions: `You are an expense processing assistant. Analyze expense reports,
    categorize them, and prepare summaries. Always verify large expenses with the user.`,
  llm: {
    provider: 'openai',
    model: 'gpt-4o',
    apiKey: { env: 'OPENAI_API_KEY' },
  },
  inputSchema: {
    description: z.string().describe('Expense description'),
    amount: z.number().describe('Expense amount in USD'),
    category: z.string().optional().describe('Optional category hint'),
  },
  tools: [ExpenseCategoryTool, ReceiptAnalyzerTool],
})
class ExpenseProcessorAgent extends AgentContext {
  async execute(input: { description: string; amount: number; category?: string }) {
    // Large expenses require manager approval before AI processing
    if (input.amount > 1000) {
      const approval = await this.elicit(
        `Expense of $${input.amount} for "${input.description}" requires manager approval.`,
        z.object({
          approved: z.boolean().describe('Approve this expense'),
          notes: z.string().optional().describe('Optional approval notes'),
        })
      );

      if (approval.status !== 'accept' || !approval.content?.approved) {
        return {
          status: 'rejected',
          reason: approval.status === 'accept' ? 'Not approved' : `User ${approval.status}d`,
        };
      }

      // Notify that approval was received
      await this.notify(`Expense approved by manager`, 'info');
    }

    // Run the default agent loop (LLM + tools) to process the expense
    return super.execute(input);
  }
}

Most agents don’t need to override execute(). The default implementation runs the LLM loop automatically. Override only when you need custom pre/post processing like approval workflows.

Elicitation Modes

FrontMCP supports two elicitation modes per the MCP specification:

Form Mode (Default)

Displays a form to collect structured input directly in the client:

const result = await this.elicit(
  'Please provide your shipping preferences',
  z.object({
    address: z.string(),
    priority: z.enum(['standard', 'express', 'overnight']),
    signature: z.boolean().default(false),
  }),
  { mode: 'form' } // Default, can be omitted
);

URL Mode

Redirects the user to an external URL for out-of-band interaction (OAuth, payments, etc.). URL mode requires an elicitationId — an opaque identifier you generate to correlate the out-of-band callback with the pending elicitation:

import { randomUUID } from '@frontmcp/utils';

const elicitationId = randomUUID();

const result = await this.elicit(
  'Please authorize access to your GitHub account',
  z.object({
    authorized: z.boolean(),
    scope: z.array(z.string()),
  }),
  {
    mode: 'url',
    elicitationId, // Required for URL mode tracking
  }
);

Security: Never put a raw session ID into the elicitation URL. Generate an opaque elicitationId (e.g. via randomUUID()) and store the mapping to the session server-side, or use a signed/encrypted payload that you can validate on callback.

Elicitation Result

The elicit() method returns an ElicitResult with the following structure:

interface ElicitResult<T> {
  status: 'accept' | 'cancel' | 'decline';
  content?: T;  // Only present when status is 'accept'
}

Status Values

Status	Description	Content Present
`accept`	User submitted the form / completed the action	Yes
`cancel`	User cancelled the elicitation	No
`decline`	User declined to provide the requested input	No

Handling Responses

const result = await this.elicit('Confirm action?', schema);

switch (result.status) {
  case 'accept':
    // User submitted - content is typed according to your schema
    console.log(result.content); // { confirmed: boolean, ... }
    break;

  case 'cancel':
    // User cancelled - no content
    return { cancelled: true };

  case 'decline':
    // User declined - no content
    return { declined: true };
}

Schema Validation

When the client returns an elicitation result with action: 'accept', FrontMCP automatically validates the content against the original schema you provided. This ensures type safety and data integrity.

How It Works

When you call this.elicit(message, schema), FrontMCP stores the schema alongside the pending elicitation record
When the user responds, FrontMCP validates the content against the stored schema
Invalid content is rejected with an InvalidInputError (same as tool input validation)

Validation Behavior

Scenario	Validation	Result
`accept` with valid content	✅ Validated	Success
`accept` with invalid content	✅ Validated	`InvalidInputError` thrown
`cancel` or `decline`	⏭️ Skipped	Success (no content expected)
No schema stored (backward compat)	⏭️ Skipped	Success

Error Handling

Invalid elicitation content throws the same InvalidInputError used for tool input validation. This allows LLMs (especially in fallback mode) to recognize the error format and retry with corrected data:

{
  "content": [{
    "type": "text",
    "text": "Invalid elicitation result content\nDetails: [{\"path\":[\"age\"],\"message\":\"Expected number, received string\"}]"
  }],
  "isError": true,
  "_meta": {
    "errorId": "err_a1b2c3d4e5f6g7h8",
    "code": "INVALID_INPUT"
  }
}

The error includes:

path: The JSON path to the invalid field (e.g., ["user", "age"])
message: Human-readable validation error message

Example: Schema Mismatch

// Tool defines a schema
const result = await this.elicit(
  'Enter your details',
  z.object({
    name: z.string(),
    age: z.number(),      // Expects number
  })
);

// If client sends { name: "John", age: "thirty" }
// FrontMCP throws InvalidInputError:
// "Invalid elicitation result content"
// Details: [{ path: ["age"], message: "Expected number, received string" }]

Schema validation uses the same InvalidInputError as tool input validation (HTTP 400, code INVALID_INPUT). This consistency helps LLMs understand and recover from validation errors.

Timeout Handling

Elicitation requests have a configurable timeout. When the timeout expires, an ElicitationTimeoutError is thrown to kill the tool execution and release resources.

Default Timeout

The default timeout is 5 minutes (300,000ms):

// Uses default 5 minute timeout
const result = await this.elicit('Confirm?', schema);

Custom Timeout

Specify a custom timeout with the ttl option:

// 1 minute timeout
const result = await this.elicit('Quick confirmation needed', schema, {
  ttl: 60000,
});

// 10 minute timeout for complex forms
const result = await this.elicit('Complete the following form', complexSchema, {
  ttl: 600000,
});

Handling Timeouts

Timeouts throw an exception to ensure the tool execution is terminated:

import { ElicitationTimeoutError } from '@frontmcp/sdk';

@Tool({
  name: 'approval-tool',
  inputSchema: { action: z.string() },
})
class ApprovalTool extends ToolContext {
  async execute({ action }) {
    try {
      const result = await this.elicit(`Approve ${action}?`, schema, {
        ttl: 60000,
      });
      // Handle response...
    } catch (error) {
      if (error instanceof ElicitationTimeoutError) {
        // Timeout occurred - tool execution will be terminated
        // You can log or perform cleanup before the error propagates
        this.logger.warn(`Elicitation timed out after ${error.ttl}ms`);
      }
      throw error; // Re-throw to terminate execution
    }
  }
}

Timeouts are designed to throw exceptions to kill tool execution. Do not catch and suppress timeout errors—this defeats the purpose of resource cleanup.

Single Elicit Per Session

Only one elicitation can be pending per session at a time. If a new elicitation is requested while one is pending, the previous one is automatically cancelled with status: 'cancel'.

// First elicit starts
const result1Promise = this.elicit('First question?', schema1);

// Second elicit cancels the first one
const result2Promise = this.elicit('Second question?', schema2);

// result1Promise resolves with { status: 'cancel' }
const result1 = await result1Promise;
console.log(result1.status); // 'cancel'

// result2Promise waits for user response
const result2 = await result2Promise;

This design ensures predictable behavior and prevents conflicting elicitation UI states.

Elicitation Hooks

FrontMCP provides flow hooks to intercept elicitation requests and results. Hooks are methods inside plugin classes that run before, after, or around specific flow stages.

Hook Types

Decorator	Description
`.Will('stage')`	Runs before the stage executes
`.Did('stage')`	Runs after the stage completes
`.Stage('stage')`	Runs as the stage (replaces default)
`.Around('stage')`	Wraps stage execution (call `next()` to continue)

ElicitationRequestHook

Intercepts the elicitation request flow before sending to the client. Stages:

Stage	Description
`parseInput`	Parse and validate the elicitation parameters
`validateRequest`	Validate mode requirements (URL mode needs elicitationId)
`generateElicitId`	Generate unique elicitation ID
`storePendingRecord`	Store pending elicitation in the store
`buildRequestParams`	Build the MCP elicitation request parameters
`finalize`	Final processing before sending

ElicitationResultHook

Intercepts the elicitation result flow when the user responds. Stages:

Stage	Description
`parseInput`	Parse the result payload
`lookupPending`	Lookup the pending elicitation record by session ID
`validateContent`	Validate content against stored schema (skipped for cancel/decline)
`buildResult`	Build the typed result object
`publishResult`	Publish result via pub/sub (for distributed deployments)
`finalize`	Final processing and cleanup

Example: Plugin with Elicitation Hooks

import { DynamicPlugin, Plugin, FlowHooksOf, FlowCtxOf } from '@frontmcp/sdk';

const ElicitationRequestHook = FlowHooksOf<'elicitation:request'>('elicitation:request');
const ElicitationResultHook = FlowHooksOf<'elicitation:result'>('elicitation:result');

@Plugin({
  name: 'elicitation-audit',
  description: 'Audit logging for elicitation flows',
})
export class ElicitationAuditPlugin extends DynamicPlugin {
  private auditLog: AuditLog;

  constructor(auditLog: AuditLog) {
    super();
    this.auditLog = auditLog;
  }

  // Runs AFTER storePendingRecord stage
  @ElicitationRequestHook.Did('storePendingRecord')
  async logElicitationStarted(ctx: FlowCtxOf<'elicitation:request'>) {
    const { elicitId, sessionId, message } = ctx.state;
    await this.auditLog.record({
      type: 'elicitation_started',
      elicitId,
      sessionId,
      message,
      timestamp: Date.now(),
    });
  }

  // Runs AFTER buildResult stage
  @ElicitationResultHook.Did('buildResult')
  async logElicitationResult(ctx: FlowCtxOf<'elicitation:result'>) {
    const { sessionId, elicitResult, pendingRecord } = ctx.state;
    if (pendingRecord && elicitResult) {
      const duration = Date.now() - pendingRecord.createdAt;
      await this.auditLog.record({
        type: 'elicitation_completed',
        elicitId: pendingRecord.elicitId,
        sessionId,
        status: elicitResult.status,
        duration,
        timestamp: Date.now(),
      });
    }
  }
}

Registering the Plugin

import { FrontMcp } from '@frontmcp/sdk';
import { ElicitationAuditPlugin } from './plugins/elicitation-audit';

@FrontMcp({
  name: 'my-app',
  plugins: [ElicitationAuditPlugin],
})
class MyApp {}

Use Cases

Input Sanitization (BEFORE buildResult):

@Plugin({ name: 'elicitation-sanitizer', description: 'Sanitize elicitation input' })
class SanitizerPlugin extends DynamicPlugin {
  @ElicitationResultHook.Will('buildResult')
  async sanitizeInput(ctx: FlowCtxOf<'elicitation:result'>) {
    // Sanitize content before it's built into result
    if (ctx.state.content) {
      ctx.state.content = this.sanitize(ctx.state.content);
    }
  }
}

Analytics (AFTER finalize):

@Plugin({ name: 'elicitation-analytics', description: 'Track elicitation completions' })
class AnalyticsPlugin extends DynamicPlugin {
  @ElicitationResultHook.Did('finalize')
  async trackCompletion(ctx: FlowCtxOf<'elicitation:result'>) {
    const { pendingRecord, elicitResult } = ctx.state;
    if (pendingRecord) {
      await analytics.track('elicitation_completed', {
        elicitId: pendingRecord.elicitId,
        duration: Date.now() - pendingRecord.createdAt,
        status: elicitResult?.status,
      });
    }
  }
}

Conditional Hooks (with filter):

@Plugin({ name: 'admin-audit', description: 'Audit admin elicitation actions' })
class AdminAuditPlugin extends DynamicPlugin {
  @ElicitationRequestHook.Did('finalize', {
    filter: (ctx) => ctx.authInfo?.role === 'admin',
  })
  async auditAdminElicitation(ctx: FlowCtxOf<'elicitation:request'>) {
    // Only runs for admin users
    await this.auditLog.recordAdminAction(ctx.state);
  }
}

Client Capability Detection

Not all MCP clients support elicitation. FrontMCP automatically checks client capabilities before sending elicitation requests.

Automatic Validation

When you call this.elicit(), FrontMCP:

Checks if the client declared elicitation capability during initialization
Verifies the client supports the requested mode (form or url)
Throws ElicitationNotSupportedError if elicitation is not supported

Manual Capability Check

You can check elicitation support before calling elicit():

import { supportsElicitation } from '@frontmcp/sdk';

@Tool({
  name: 'conditional-elicit',
  inputSchema: { action: z.string() },
})
class ConditionalElicitTool extends ToolContext {
  async execute({ action }) {
    const capabilities = this.scope.notifications.getClientCapabilities(
      this.context.authInfo?.sessionId
    );

    if (supportsElicitation(capabilities, 'form')) {
      // Client supports form elicitation
      const result = await this.elicit('Confirm?', schema);
      // ...
    } else {
      // Fallback for clients without elicitation support
      return { warning: 'This action requires user confirmation which is not supported by your client' };
    }
  }
}

Capability Helper

The supportsElicitation helper function:

import { supportsElicitation, ClientCapabilities } from '@frontmcp/sdk';

// Check for any elicitation support
supportsElicitation(capabilities); // true/false

// Check for specific mode
supportsElicitation(capabilities, 'form'); // true/false
supportsElicitation(capabilities, 'url');  // true/false

Universal LLM Support

FrontMCP elicitation works with all LLMs, not just those that support the MCP elicitation protocol. For clients that don’t support elicitation (like OpenAI, Google Gemini, Cursor, etc.), FrontMCP automatically falls back to an LLM-mediated approach.

Zero code changes required! Your tools use the same this.elicit() API regardless of client capabilities. FrontMCP handles detection and routing automatically.

How It Works

Standard Elicitation (Claude, supporting clients): Fallback Elicitation (OpenAI, Gemini, etc.):

Client Compatibility

Client	Elicitation Support	Mechanism
Claude Desktop	✅ Native	Standard MCP protocol
Claude.ai	✅ Native	Standard MCP protocol
OpenAI ChatGPT	✅ Automatic	Fallback via `sendElicitationResult`
Google Gemini	✅ Automatic	Fallback via `sendElicitationResult`
Cursor	✅ Automatic	Fallback via `sendElicitationResult`
Custom Clients	Depends	Check `experimental.elicitation` capability

Fallback Response Format

When a tool calls elicit() and the client doesn’t support elicitation, FrontMCP returns a structured response:

{
  "content": [
    {
      "type": "text",
      "text": "This tool requires user input to continue.\n\n**Question:** Are you sure?\n\n..."
    }
  ],
  "_meta": {
    "elicitationPending": {
      "elicitId": "elicit-1234-abc",
      "message": "Are you sure?",
      "schema": { "type": "object", "properties": { "confirmed": { "type": "boolean" } } },
      "instructions": "Call sendElicitationResult tool after collecting user input"
    }
  }
}

The LLM reads these instructions, asks the user, then calls the sendElicitationResult tool:

{
  "name": "sendElicitationResult",
  "arguments": {
    "elicitId": "elicit-1234-abc",
    "action": "accept",
    "content": { "confirmed": true }
  }
}

The sendElicitationResult Tool

This system tool is automatically registered for clients that don’t support elicitation. It:

Is hidden from clients that support standard elicitation
Accepts the elicitId, user action, and content
Re-invokes the original tool with the result pre-injected
Returns the original tool’s final result

Input Schema:

Parameter	Type	Description
`elicitId`	`string`	The elicitation ID from the pending request
`action`	`'accept' \| 'cancel' \| 'decline'`	User’s action
`content`	`unknown`	User’s response (required when action is `'accept'`)

Key Benefits

Transparent to developers — Same this.elicit() API for all clients
Automatic detection — Framework checks client capabilities at runtime
Works with Redis — Fallback state is stored in Redis for distributed deployments
Type-safe — Schema is converted to JSON Schema for LLM understanding

Error Handling

ElicitationNotSupportedError

Thrown when:

Client doesn’t support elicitation
Client doesn’t support the requested mode (form/url)
No session is available
Transport is not available

import { ElicitationNotSupportedError } from '@frontmcp/sdk';

try {
  const result = await this.elicit('Question?', schema);
} catch (error) {
  if (error instanceof ElicitationNotSupportedError) {
    // Handle gracefully - perhaps skip the confirmation
    return { skippedConfirmation: true, reason: error.message };
  }
  throw error;
}

ElicitationTimeoutError

Thrown when the user doesn’t respond within the TTL:

import { ElicitationTimeoutError } from '@frontmcp/sdk';

try {
  const result = await this.elicit('Question?', schema, { ttl: 30000 });
} catch (error) {
  if (error instanceof ElicitationTimeoutError) {
    console.log(`Timed out after ${error.ttl}ms`);
    console.log(`Elicit ID: ${error.elicitId}`);
  }
  throw error;
}

Configuration Options

The elicit() method accepts an options object:

interface ElicitOptions {
  mode?: 'form' | 'url';    // Default: 'form'
  ttl?: number;              // Default: 300000 (5 minutes)
  elicitationId?: string;    // Required for URL mode
}

Option	Description	Default
`mode`	Elicitation mode (`'form'` or `'url'`)	`'form'`
`ttl`	Timeout in milliseconds before throwing	300000
`elicitationId`	Unique ID for URL mode correlation	Auto (form) / Required (url)

Distributed Deployments

By default, elicitation state is stored in-memory on the server that initiated the elicit request. This works for single-node deployments but fails in distributed environments where different nodes may handle the initial request and the user’s response.

The Problem

In a distributed deployment:

Redis Mode (Recommended for Production)

When Redis is configured, FrontMCP automatically uses Redis for elicitation state storage and pub/sub for cross-node result routing:

@FrontMcp({
  // Configure Redis for distributed elicitation
  redis: {
    provider: 'redis',
    host: process.env.REDIS_HOST ?? 'localhost',
    port: parseInt(process.env.REDIS_PORT ?? '6379'),
    password: process.env.REDIS_PASSWORD,
  },
})
class MyApp {}

With Redis configured, elicitation works seamlessly across nodes:

Single-Node Mode (Development)

Without Redis, FrontMCP uses in-memory storage with a warning:

[FrontMCP] Elicitation: using in-memory store (single-node mode).
Configure Redis for distributed deployments.

This is appropriate for:

Local development
Single-node deployments
Testing environments

Sticky Sessions (Alternative)

If you cannot use Redis, configure your load balancer for session affinity to ensure requests from the same client always reach the same server: Nginx:

upstream backend {
  ip_hash;  # Sticky sessions by client IP
  server node1:3000;
  server node2:3000;
}

AWS ALB:

Enable “Stickiness” in target group settings
Use application-based cookie (recommended) or duration-based stickiness
Set appropriate stickiness duration (longer than your elicitation TTL)

Kubernetes (Ingress):

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  annotations:
    nginx.ingress.kubernetes.io/affinity: "cookie"
    nginx.ingress.kubernetes.io/session-cookie-name: "mcp-session"
    nginx.ingress.kubernetes.io/session-cookie-expires: "172800"

Sticky sessions are less reliable than Redis because they depend on load balancer behavior and can fail during node restarts or scaling events. Use Redis for production deployments.

Vercel Edge Functions

Edge functions are stateless and require Redis for elicitation. An error is thrown if elicitation is attempted without Redis on Edge:

Error: Elicitation requires Redis configuration when running on Edge runtime.
Edge functions are stateless and cannot use in-memory elicitation.
Configure redis in @FrontMcp({ redis: { provider: "redis", host: "...", port: ... } })

Configure Redis when deploying to Vercel Edge:

@FrontMcp({
  redis: {
    provider: 'redis',
    host: process.env.UPSTASH_REDIS_HOST!,
    port: parseInt(process.env.UPSTASH_REDIS_PORT ?? '6379'),
    password: process.env.UPSTASH_REDIS_PASSWORD,
    tls: true,  // Required for Upstash
  },
})
class MyApp {}

Upstash provides serverless Redis that works well with Vercel Edge functions.

Deployment Mode Summary

Environment	Storage	Notes
Local development	In-memory	Automatic, no configuration needed
Single-node production	In-memory	Works but shows warning
Multi-node production	Redis	Configure `redis` in `@FrontMcp`
Vercel Edge	Redis (required)	Use Upstash or similar
Kubernetes	Redis	Use Redis cluster for HA
Load balanced (no Redis)	Sticky sessions	Configure session affinity

Encrypted Elicitation

When collecting sensitive user data (passwords, PII, payment information), FrontMCP can encrypt all elicitation data at rest using session-derived keys.

Why Encrypt?

Elicitation often collects sensitive data:

Account credentials and API keys
Personal information (SSN, address, phone)
Financial data (card numbers, bank accounts)
Health information (HIPAA-protected data)

Without encryption, this data is stored in plaintext in Redis or memory, accessible to anyone with database access.

Security Model

Encrypted elicitation provides zero-knowledge storage:

Session-Derived Keys: Each session gets a unique encryption key derived from HKDF-SHA256(serverSecret + sessionId)
AES-256-GCM: Authenticated encryption prevents tampering
Cross-Session Isolation: Session B cannot decrypt Session A’s data, even with database access
Fail-Safe Decryption: Decryption fails silently for tampered or corrupted data

Configuration

Set one of these environment variables to enable encryption:

# Recommended - dedicated secret for elicitation
MCP_ELICITATION_SECRET=your-secure-random-secret-at-least-32-chars

# Fallback options (checked in order)
MCP_SESSION_SECRET=your-session-secret
MCP_SERVER_SECRET=your-server-secret

Encryption is automatic when a secret is available. No code changes required.

How It Works

Cross-Session Isolation

Even if an attacker gains database access, they cannot decrypt elicitation data without:

The server secret (MCP_ELICITATION_SECRET)
The specific sessionId used to encrypt the data

// Session A encrypts sensitive data
const resultA = await this.elicit('Enter your API key', z.object({
  apiKey: z.string(),
}));
// Stored as encrypted blob with Session A's derived key

// Session B cannot decrypt Session A's data
// Even with database access, decryption returns null

Best Practices

Do:

Use a strong random secret (32+ characters)
Rotate secrets periodically (with care for in-flight elicitations)
Set secrets via environment variables, not code
Use MCP_ELICITATION_SECRET for dedicated elicitation security

Don’t:

Store secrets in version control
Use weak or predictable secrets
Share secrets across environments (dev/staging/prod)
Log or expose encrypted data for debugging

Real-World Examples

Approval Workflow

@Tool({
  name: 'approve-expense',
  description: 'Process an expense report with manager approval',
  inputSchema: {
    expenseId: z.string(),
    amount: z.number(),
    description: z.string(),
  },
  annotations: {
    destructiveHint: true,
  },
})
class ApproveExpenseTool extends ToolContext {
  async execute({ expenseId, amount, description }) {
    // Large expenses require approval
    if (amount > 500) {
      const approval = await this.elicit(
        `Expense #${expenseId} for $${amount.toFixed(2)}: "${description}"`,
        z.object({
          approved: z.boolean().describe('Approve this expense'),
          notes: z.string().optional().describe('Optional approval notes'),
        }),
        { ttl: 300000 } // 5 minute timeout
      );

      if (approval.status !== 'accept') {
        return {
          status: 'cancelled',
          expenseId,
          reason: `Manager ${approval.status}d the approval request`
        };
      }

      if (!approval.content?.approved) {
        return {
          status: 'rejected',
          expenseId,
          notes: approval.content?.notes,
        };
      }
    }

    // Process the expense
    await this.processExpense(expenseId);
    return { status: 'approved', expenseId };
  }
}

Multi-Step Confirmation

@Tool({
  name: 'transfer-funds',
  description: 'Transfer funds between accounts',
  inputSchema: {
    fromAccount: z.string(),
    toAccount: z.string(),
    amount: z.number(),
  },
})
class TransferFundsTool extends ToolContext {
  async execute({ fromAccount, toAccount, amount }) {
    // Step 1: Confirm transfer details
    const confirm = await this.elicit(
      `Transfer $${amount} from ${fromAccount} to ${toAccount}?`,
      z.object({
        confirmed: z.boolean(),
      })
    );

    if (confirm.status !== 'accept' || !confirm.content?.confirmed) {
      return { cancelled: true };
    }

    // Step 2: Verify with additional security
    const verification = await this.elicit(
      'Please verify this transfer',
      z.object({
        verificationCode: z.string().min(6).max(6).describe('Enter the 6-digit code sent to your phone'),
      }),
      { ttl: 120000 } // 2 minute timeout for verification
    );

    if (verification.status !== 'accept') {
      return { cancelled: true, reason: 'Verification not completed' };
    }

    // Validate code and process transfer
    if (!this.validateCode(verification.content?.verificationCode)) {
      return { error: 'Invalid verification code' };
    }

    await this.executeTransfer(fromAccount, toAccount, amount);
    return { success: true, transactionId: this.generateTxId() };
  }
}

OAuth Authorization (URL Mode)

import { randomUUID } from '@frontmcp/utils';

@Tool({
  name: 'connect-github',
  description: 'Connect your GitHub account',
  inputSchema: {},
})
class ConnectGitHubTool extends ToolContext {
  async execute() {
    // Generate an opaque correlation id for the OAuth callback
    const elicitationId = randomUUID();

    // Server-side mapping from correlation id to session
    await this.get(OAuthStateStore).set(elicitationId, {
      provider: 'github',
      sessionId: this.context.authInfo?.sessionId,
      startedAt: Date.now(),
    });

    const result = await this.elicit(
      'Click the link to authorize FrontMCP to access your GitHub account',
      z.object({
        accessToken: z.string(),
        scope: z.array(z.string()),
        username: z.string(),
      }),
      {
        mode: 'url',
        elicitationId,
        ttl: 300000,
      }
    );

    if (result.status !== 'accept') {
      return { connected: false, reason: `Authorization ${result.status}d` };
    }

    // Store the token and return success
    await this.get(TokenStore).save(
      this.context.authInfo?.sessionId,
      'github',
      result.content!.accessToken
    );

    return {
      connected: true,
      username: result.content!.username,
      scope: result.content!.scope,
    };
  }
}

The elicitationId is an opaque correlation handle. Map it to the session in your own server-side store so the OAuth callback can resolve the session without exposing raw session IDs in URLs.

Graceful Fallback

@Tool({
  name: 'delete-file',
  description: 'Delete a file with optional confirmation',
  inputSchema: {
    path: z.string(),
    skipConfirmation: z.boolean().optional(),
  },
})
class DeleteFileTool extends ToolContext {
  async execute({ path, skipConfirmation }) {
    // Check if we should/can request confirmation
    const shouldConfirm = !skipConfirmation;
    const capabilities = this.scope.notifications.getClientCapabilities(
      this.context.authInfo?.sessionId
    );
    const canElicit = supportsElicitation(capabilities, 'form');

    if (shouldConfirm && canElicit) {
      const result = await this.elicit(
        `Delete "${path}"? This cannot be undone.`,
        z.object({ confirmed: z.boolean() })
      );

      if (result.status !== 'accept' || !result.content?.confirmed) {
        return { deleted: false, reason: 'User did not confirm' };
      }
    } else if (shouldConfirm && !canElicit) {
      // Client doesn't support elicitation - return warning
      return {
        deleted: false,
        reason: 'Confirmation required but client does not support elicitation. Use skipConfirmation: true to proceed.',
      };
    }

    // Proceed with deletion
    await this.deleteFile(path);
    return { deleted: true, path };
  }
}

MCP Protocol Integration

Elicitation integrates with the MCP protocol via:

Flow	Description
`elicitation/create`	Server sends elicitation request to client
Result callback	Client sends result with action and content

Client Capabilities

Clients advertise elicitation support during MCP initialization:

{
  "capabilities": {
    "elicitation": {
      "form": {},
      "url": {}
    }
  }
}

FrontMCP automatically reads these capabilities and validates them before sending elicitation requests.

Protocol Flow

Tool calls this.elicit(message, schema, options)
FrontMCP validates client capabilities
FrontMCP sends elicitation/create request to client
Client displays form/redirects user based on mode
User interacts with the elicitation UI
Client sends result back with action (‘accept’, ‘cancel’, ‘decline’) and optional content
FrontMCP resolves the promise with typed ElicitResult

For the full protocol specification, see MCP Elicitation.

Best Practices

Do:

Use descriptive messages that clearly explain what information is needed
Define clear Zod schemas with .describe() on each field
Set appropriate timeouts based on the complexity of the request
Handle all three status values (accept, cancel, decline)
Provide graceful fallbacks for clients without elicitation support
Use form mode for simple confirmations and structured data collection
Use URL mode only for external authentication flows

Don’t:

Suppress timeout errors—let them propagate to terminate execution
Request elicitation for every tool—only when user input is truly needed
Set extremely long timeouts that could leave resources hanging
Assume all clients support elicitation—always check or handle errors
Chain multiple dependent elicitations without clear user guidance
Use elicitation for data that should be provided upfront in tool input

API Reference

Types

// Status of an elicitation result
type ElicitStatus = 'accept' | 'cancel' | 'decline';

// Result returned by elicit()
interface ElicitResult<T = unknown> {
  status: ElicitStatus;
  content?: T;  // Only present when status is 'accept'
}

// Options for elicit()
interface ElicitOptions {
  mode?: 'form' | 'url';
  ttl?: number;
  elicitationId?: string;
}

// Default timeout constant
const DEFAULT_ELICIT_TTL = 300000; // 5 minutes

Errors

// Thrown when client doesn't support elicitation
class ElicitationNotSupportedError extends PublicMcpError {
  // code: 'ELICITATION_NOT_SUPPORTED'
  // statusCode: 400
}

// Thrown when elicitation times out
class ElicitationTimeoutError extends PublicMcpError {
  readonly elicitId: string;
  readonly ttl: number;
  // code: 'ELICITATION_TIMEOUT'
  // statusCode: 408
}

Helpers

// Check if client supports elicitation
function supportsElicitation(
  capabilities?: ClientCapabilities,
  mode?: 'form' | 'url'
): boolean;

​Why Elicitation?

​Basic Usage

​Class Style (ToolContext)

​Class Style (AgentContext)

​Elicitation Modes

​Form Mode (Default)

​URL Mode

​Elicitation Result

​Status Values

​Handling Responses

​Schema Validation

​How It Works

​Validation Behavior

​Error Handling

​Example: Schema Mismatch

​Timeout Handling

​Default Timeout

​Custom Timeout

​Handling Timeouts

​Single Elicit Per Session

​Elicitation Hooks

​Hook Types

​ElicitationRequestHook

​ElicitationResultHook

​Example: Plugin with Elicitation Hooks

​Registering the Plugin

​Use Cases

​Client Capability Detection

​Automatic Validation

​Manual Capability Check

​Capability Helper

​Universal LLM Support

​How It Works

​Client Compatibility

​Fallback Response Format

​The sendElicitationResult Tool

​Key Benefits

​Error Handling

​ElicitationNotSupportedError

​ElicitationTimeoutError

​Configuration Options

​Distributed Deployments

​The Problem

​Redis Mode (Recommended for Production)

​Single-Node Mode (Development)

​Sticky Sessions (Alternative)

​Vercel Edge Functions

​Deployment Mode Summary

​Encrypted Elicitation

​Why Encrypt?

​Security Model

​Configuration

​How It Works

​Cross-Session Isolation

​Best Practices

​Real-World Examples

​Approval Workflow

​Multi-Step Confirmation

​OAuth Authorization (URL Mode)

​Graceful Fallback

​MCP Protocol Integration

​Client Capabilities

​Protocol Flow

​Best Practices

​API Reference

​Types

​Errors

​Helpers

Why Elicitation?

Basic Usage

Class Style (ToolContext)

Class Style (AgentContext)

Elicitation Modes

Form Mode (Default)

URL Mode

Elicitation Result

Status Values

Handling Responses

Schema Validation

How It Works

Validation Behavior

Error Handling

Example: Schema Mismatch

Timeout Handling

Default Timeout

Custom Timeout

Handling Timeouts

Single Elicit Per Session

Elicitation Hooks

Hook Types

ElicitationRequestHook

ElicitationResultHook

Example: Plugin with Elicitation Hooks

Registering the Plugin

Use Cases

Client Capability Detection

Automatic Validation

Manual Capability Check

Capability Helper

Universal LLM Support

How It Works

Client Compatibility

Fallback Response Format

The sendElicitationResult Tool

Key Benefits

Error Handling

ElicitationNotSupportedError

ElicitationTimeoutError

Configuration Options

Distributed Deployments

The Problem

Redis Mode (Recommended for Production)

Single-Node Mode (Development)

Sticky Sessions (Alternative)

Vercel Edge Functions

Deployment Mode Summary

Encrypted Elicitation

Why Encrypt?

Security Model

Configuration

How It Works

Cross-Session Isolation

Best Practices

Real-World Examples

Approval Workflow

Multi-Step Confirmation

OAuth Authorization (URL Mode)

Graceful Fallback

MCP Protocol Integration

Client Capabilities

Protocol Flow

Best Practices

API Reference

Types

Errors

Helpers