Client Guidance

Client Guidance

Purpose

Design MCP error responses that guide clients toward success through progressive enhancement: suggesting similar tools, recommending correct parameters, providing schema subsets, and offering corrective feedback rather than just rejecting requests.

When to Use

• Designing error handling for MCP tools
• Implementing "did you mean" suggestions
• Adding fuzzy matching for tool/parameter names
• Including schema hints in error responses
• Building robust client guidance patterns

Core Principle: Errors as Progressive Enhancement

Traditional errors reject and stop. Progressive enhancement errors reject but guide:

Traditional:
  → "Unknown tool: serach"
  → Client is stuck

Progressive Enhancement:
  → "Unknown tool: serach"
  → "Did you mean: search, search_code, search_files?"
  → "Available tool categories: search (3), get (5), list (4)"
  → Client can self-correct

Pattern 1: Similar Tool Suggestions

When a client calls an invalid tool, suggest similar ones:

Implementation

// Pseudocode
function handleToolCall(toolName: string, params: any) {
  const tool = tools.get(toolName)

  if (!tool) {
    const similar = findSimilarTools(toolName, tools.keys())
    const categories = groupToolsByCategory(tools.keys())

    return {
      error: {
        code: "UNKNOWN_TOOL",
        message: `Tool '${toolName}' not found`,
        suggestions: {
          similar_tools: similar.slice(0, 5),
          did_you_mean: similar[0],
          available_categories: Object.keys(categories),
          hint: similar.length > 0
            ? `Try: ${similar[0]}`
            : "Use info() to list all available tools"
        }
      }
    }
  }

  return tool.execute(params)
}

function findSimilarTools(input: string, available: string[]): string[] {
  // Levenshtein distance + prefix matching + substring matching
  return available
    .map(name => ({
      name,
      score: calculateSimilarity(input, name)
    }))
    .filter(t => t.score > 0.3)
    .sort((a, b) => b.score - a.score)
    .map(t => t.name)
}

Example Response

Request: serach(pattern: "User")

{
  "error": {
    "code": "UNKNOWN_TOOL",
    "message": "Tool 'serach' not found",
    "suggestions": {
      "did_you_mean": "search",
      "similar_tools": ["search", "search_code", "search_files"],
      "available_categories": ["search", "get", "list", "info"],
      "hint": "Try: search(pattern: \"User\")"
    }
  }
}

Similarity Algorithm

Combine multiple signals for best matches:

function calculateSimilarity(input: string, candidate: string): number {
  const scores = [
    levenshteinSimilarity(input, candidate) * 0.4,   // Typo tolerance
    prefixMatch(input, candidate) * 0.3,              // Starts with
    substringMatch(input, candidate) * 0.2,           // Contains
    wordOverlap(input, candidate) * 0.1               // Semantic similarity
  ]
  return scores.reduce((a, b) => a + b, 0)
}

Priority order for suggestions:

1. Exact prefix match (sea → search)
2. Low edit distance (serach → search)
3. Contains substring (auth → authenticate)
4. Word overlap (user_search → search_users)

Pattern 2: Parameter Correction

When parameters are invalid, suggest corrections:

Unknown Parameter Handling

function validateParams(params: any, schema: Schema) {
  const known = Object.keys(schema.properties)
  const provided = Object.keys(params)
  const unknown = provided.filter(p => !known.includes(p))

  const warnings = []
  const suggestions = {}

  for (const param of unknown) {
    const similar = findSimilarParams(param, known)
    if (similar.length > 0) {
      suggestions[param] = {
        did_you_mean: similar[0],
        alternatives: similar.slice(0, 3)
      }
      warnings.push(`Unknown param '${param}', did you mean '${similar[0]}'?`)
    } else {
      warnings.push(`Unknown param '${param}' ignored`)
    }
  }

  return { warnings, suggestions, valid: true }
}

Example Response

Request: search(patern: "User", filtr: "*.ts")

{
  "results": [...],
  "warnings": [
    "Unknown param 'patern', did you mean 'pattern'?",
    "Unknown param 'filtr', did you mean 'filter'?"
  ],
  "suggestions": {
    "patern": {
      "did_you_mean": "pattern",
      "alternatives": ["pattern"]
    },
    "filtr": {
      "did_you_mean": "filter",
      "alternatives": ["filter", "file_filter"]
    }
  },
  "hint": "Corrected call: search(pattern: \"User\", filter: \"*.ts\")"
}

Auto-Correction Option

For high-confidence matches, optionally auto-correct:

function handleParams(params: any, schema: Schema) {
  const corrected = {}
  const corrections = []

  for (const [key, value] of Object.entries(params)) {
    if (schema.properties[key]) {
      corrected[key] = value
    } else {
      const match = findBestMatch(key, Object.keys(schema.properties))
      if (match && match.confidence > 0.9) {
        // Auto-correct with notification
        corrected[match.name] = value
        corrections.push({
          from: key,
          to: match.name,
          confidence: match.confidence,
          auto_corrected: true
        })
      }
    }
  }

  return { corrected, corrections }
}

Response with auto-correction:

{
  "results": [...],
  "corrections": [
    {
      "from": "patern",
      "to": "pattern",
      "confidence": 0.95,
      "auto_corrected": true
    }
  ],
  "note": "Parameter 'patern' was auto-corrected to 'pattern'"
}

Pattern 3: Schema Hints in Errors

Include relevant schema information in error responses:

Missing Required Parameter

{
  "error": {
    "code": "MISSING_REQUIRED",
    "message": "Required parameter 'pattern' is missing",
    "schema_hint": {
      "required": ["pattern"],
      "optional": ["filter", "max", "offset"],
      "example": {
        "pattern": "authenticate",
        "filter": "*.ts",
        "max": 50
      }
    },
    "documentation_url": "https://docs.example.com/tools/search"
  }
}

Invalid Type

{
  "error": {
    "code": "INVALID_TYPE",
    "message": "Parameter 'max' must be integer, got string",
    "details": {
      "parameter": "max",
      "expected_type": "integer",
      "received_type": "string",
      "received_value": "fifty"
    },
    "schema_hint": {
      "max": {
        "type": "integer",
        "minimum": 1,
        "maximum": 1000,
        "default": 50,
        "description": "Maximum number of results to return"
      }
    },
    "hint": "Use a number: max: 50"
  }
}

Out of Range

{
  "error": {
    "code": "OUT_OF_RANGE",
    "message": "Parameter 'max' value 5000 exceeds maximum 1000",
    "details": {
      "parameter": "max",
      "value": 5000,
      "minimum": 1,
      "maximum": 1000
    },
    "hint": "Use max: 1000 for maximum results, or paginate with offset"
  }
}

Pattern 4: Progressive Schema Disclosure

Provide schema at different detail levels based on context:

Level 1: Minimal (in error messages)

{
  "error": {
    "code": "INVALID_INPUT",
    "schema_hint": {
      "required": ["pattern"],
      "optional": ["filter", "max"]
    }
  }
}

Level 2: With Types (on request)

{
  "schema": {
    "pattern": {"type": "string", "required": true},
    "filter": {"type": "string", "required": false},
    "max": {"type": "integer", "required": false, "default": 50}
  }
}

Level 3: Full (via info tool or --schema flag)

{
  "schema": {
    "pattern": {
      "type": "string",
      "required": true,
      "description": "Search pattern (regex supported)",
      "examples": ["authenticate", "User.*", "^function"]
    },
    "filter": {
      "type": "string",
      "required": false,
      "description": "File filter glob pattern",
      "examples": ["*.ts", "src/**/*.js"]
    },
    "max": {
      "type": "integer",
      "required": false,
      "default": 50,
      "minimum": 1,
      "maximum": 1000,
      "description": "Maximum results to return"
    }
  }
}

Implementation

function getSchemaHint(schema: Schema, level: 'minimal' | 'typed' | 'full') {
  switch (level) {
    case 'minimal':
      return {
        required: schema.required,
        optional: Object.keys(schema.properties).filter(k => !schema.required.includes(k))
      }
    case 'typed':
      return Object.fromEntries(
        Object.entries(schema.properties).map(([k, v]) => [k, {
          type: v.type,
          required: schema.required.includes(k),
          default: v.default
        }])
      )
    case 'full':
      return schema.properties
  }
}

Pattern 5: Contextual Next Steps

Always provide actionable next steps:

After Successful Query

{
  "results": [...],
  "has_more": true,
  "total": 127,
  "next_steps": {
    "get_more": "search(pattern: \"User\", offset: 10)",
    "get_details": "get_definition(id: \"r1\")",
    "refine_search": "search(pattern: \"User\", filter: \"*.ts\")"
  }
}

After Error

{
  "error": {
    "code": "NOT_FOUND",
    "message": "No results found for pattern 'xyzabc123'",
    "next_steps": {
      "broaden_search": "Try a simpler pattern or remove filters",
      "check_scope": "Use info() to see indexed paths",
      "alternatives": [
        "search(pattern: \"xyz\")",
        "search(pattern: \"abc\")",
        "list_files()"
      ]
    }
  }
}

After Partial Success

{
  "results": [...],
  "warnings": ["Some paths were not accessible"],
  "partial": true,
  "next_steps": {
    "retry_failed": "search(pattern: \"User\", include_errors: true)",
    "check_permissions": "info(detail: \"permissions\")",
    "proceed_anyway": "Results shown are valid, continue with get_definition(id)"
  }
}

Pattern 6: Preemptive Guidance

Include helpful hints even in successful responses:

Query Tool Response

{
  "results": [
    {"id": "r1", "name": "User.authenticate", "confidence": 0.95},
    {"id": "r2", "name": "Auth.validate", "confidence": 0.72}
  ],
  "has_more": true,
  "total": 47,
  "tips": {
    "high_results": "Use filter: \"*.ts\" to narrow to TypeScript files",
    "get_details": "Use get_definition(id) for full code and documentation",
    "pagination": "Use offset: 10 to see more results"
  }
}

When Approaching Limits

{
  "results": [...],
  "returned": 100,
  "total": 100,
  "limit_reached": true,
  "guidance": {
    "at_limit": "Showing maximum 100 results",
    "recommendations": [
      "Add filter to narrow results: filter: \"src/**\"",
      "Use more specific pattern",
      "Results are sorted by relevance, best matches shown first"
    ]
  }
}

Error Severity and Response Strategy

Severity Levels

Severity	Action	Include Schema	Suggest Similar
Critical	Reject	Full schema	Yes
High	Reject	Typed schema	Yes
Medium	Warn + Continue	Minimal	Yes
Low	Warn only	No	Optional

Decision Matrix

function determineErrorSeverity(error: ValidationError): Severity {
  // Critical: Security risks, data corruption potential
  if (error.type === 'injection_attempt') return 'critical'
  if (error.type === 'path_traversal') return 'critical'

  // High: Required field missing, type mismatch
  if (error.type === 'missing_required') return 'high'
  if (error.type === 'type_mismatch') return 'high'

  // Medium: Unknown params, out of range (correctable)
  if (error.type === 'unknown_param') return 'medium'
  if (error.type === 'out_of_range') return 'medium'

  // Low: Deprecation warnings, style issues
  if (error.type === 'deprecated_param') return 'low'

  return 'medium'
}

Complete Error Response Template

{
  "error": {
    "code": "ERROR_CODE",
    "message": "Human-readable error message",
    "severity": "high|medium|low",
    "details": {
      "field": "affected_field",
      "value": "provided_value",
      "expected": "what was expected"
    },
    "suggestions": {
      "did_you_mean": "closest_match",
      "alternatives": ["option1", "option2"],
      "auto_correctable": true
    },
    "schema_hint": {
      "required": ["field1"],
      "optional": ["field2"],
      "types": {"field1": "string", "field2": "integer"}
    },
    "next_steps": {
      "fix": "Corrected call example",
      "help": "info(tool: \"tool_name\")",
      "docs": "https://docs.example.com/tool"
    }
  }
}

Implementation Checklist

For every MCP tool:

• Unknown tool calls suggest similar tools
• Unknown parameters suggest similar parameters
• Missing required fields show schema hint
• Type errors include expected vs received
• Out of range errors show valid range
• All errors include next_steps
• High-confidence typos offer auto-correction
• Schema hints match error severity
• Successful responses include tips when relevant
• Warnings are actionable, not just informational

Testing Client Guidance

Use the mcp-fuzzer agent to validate:

## Test: Unknown Tool Handling

**Input:** Call non-existent tool "serach"
**Expected:**
- Error code: UNKNOWN_TOOL
- did_you_mean present
- similar_tools array with "search"
- hint for correction

## Test: Unknown Parameter Handling

**Input:** search(patern: "User")
**Expected:**
- Warnings array present
- Suggestion for "pattern"
- Results still returned (graceful handling)

## Test: Schema Hints on Missing Required

**Input:** search() with no parameters
**Expected:**
- Error code: MISSING_REQUIRED
- schema_hint.required includes "pattern"
- Example usage provided

Quick Reference

Client guidance principles:

1. Never just reject - Always suggest alternatives
2. Fuzzy match everything - Tools, params, values
3. Progressive schema - More detail for bigger errors
4. Auto-correct when safe - High confidence typos
5. Next steps always - What should client do now?
6. Preemptive tips - Help before they need it

Error response must include:

• code - Machine-readable error code
• message - Human-readable explanation
• suggestions - Did you mean / alternatives
• next_steps - How to proceed

Similarity thresholds:

• Auto-correct: >0.9 confidence
• Suggest first: >0.7 confidence
• Show in list: >0.4 confidence
• Omit: <0.4 confidence