demo · v149

Error Handling Lab

WebMCP tool calls can fail in five distinct ways: validation error (wrong input types), runtime error (tool execution failed), timeout (response took too long), permission denied (user revoked consent), and partial success (some outputs missing). Each failure has a structured error response and a recommended recovery strategy. Select a scenario to see the full call/response cycle and how a well-implemented agent handles it.

Select an error scenario

🚫Validation Error

Wrong input schema

code: "INVALID_PARAMS"

💥Runtime Error

Tool execution failed

code: "TOOL_EXECUTION_ERROR"

⏱Timeout

Response took too long

code: "TIMEOUT"

🔒Permission Denied

User revoked tool consent

code: "PERMISSION_DENIED"

▶Partial Success

Some outputs missing

code: "PARTIAL_RESULT"

❓Tool Not Found

Capability removed from page

code: "TOOL_NOT_FOUND"

Scenario

Live simulation

// WebMCP: handle all error types from a tool call
try {
  const result = await navigator.modelContext.callTool({
    name: 'search_products',
    input: { query: 'running shoes', maxResults: 5 }
  });
  // Handle partial success
  if (result.partial) {
    console.warn('Partial result:', result.missingFields);
  }
  renderResults(result.output);
} catch (err) {
  switch (err.code) {
    case 'INVALID_PARAMS':
      // Fix the input and retry once
      break;
    case 'PERMISSION_DENIED':
      // Show UI to re-request consent
      break;
    case 'TIMEOUT':
      // Retry with exponential backoff
      break;
    case 'TOOL_NOT_FOUND':
      // Refresh tool registry before retrying
      break;
    default:
      // Surface to user, do not retry blindly
  }
}

implementation reference

Need the exact API surface, compatibility boundaries, errors, lifecycle, and source links? Read the matching gendn reference ↗

see also

implementation reference