Hugging Face MCP Server

import type { CallToolResult } from '@modelcontextprotocol/sdk/types.js'; import { type ServerNotification, type ServerRequest } from '@modelcontextprotocol/sdk/types.js'; import type { RequestHandlerExtra } from '@modelcontextprotocol/sdk/shared/protocol.js'; import { callGradioToolWithHeaders } from '@llmindset/hf-mcp'; import { logger } from './logger.js'; import { stripImageContentFromResult, extractUrlFromContent } from './gradio-result-processor.js'; import { gradioMetrics, getMetricsSafeName } from './gradio-metrics.js'; /** * Options for calling a Gradio tool */ export interface GradioToolCallOptions { /** Whether to strip image content from the result */ stripImageContent?: boolean; /** Original tool name (for logging) */ toolName: string; /** Outward-facing tool name (for logging) */ outwardFacingName: string; /** Session information for client-specific handling */ sessionInfo?: { clientSessionId?: string; isAuthenticated?: boolean; clientInfo?: { name: string; version: string }; }; /** Gradio widget URI for OpenAI client */ gradioWidgetUri?: string; /** Space name for structured content */ spaceName?: string; } /** * Unified Gradio tool caller that handles: * - SSE connection management * - MCP tool invocation * - Progress notification relay * * Returns the raw MCP result without post-processing. Callers should apply * image filtering and OpenAI-specific transforms as needed using applyResultPostProcessing. * * This ensures both proxied gr_* tools and the space tool's invoke operation * behave identically. */ export async function callGradioTool( sseUrl: string, toolName: string, parameters: Record<string, unknown>, hfToken: string | undefined, extra: RequestHandlerExtra<ServerRequest, ServerNotification> | undefined ): Promise<CallToolResult> { logger.info({ tool: toolName, params: parameters }, 'Calling Gradio tool via unified caller'); const metricsToolName = getMetricsSafeName(toolName); // Call the remote tool via shared helper (handles SSE, progress relay, header capture) const { result, capturedHeaders } = await callGradioToolWithHeaders( sseUrl, toolName, parameters, hfToken, extra, { logProxiedReplica: true, onProgressRelayFailure: () => gradioMetrics.recordProgressRelayFailure(metricsToolName), } ); // Attach captured headers (e.g., X-Proxied-Replica) to the result meta so callers can inspect them const proxiedReplica = capturedHeaders['x-proxied-replica']; if (proxiedReplica) { logger.debug({ tool: toolName, proxiedReplica }, 'Captured Gradio response header'); return { ...result, _meta: { ...(result as { _meta?: Record<string, unknown> })._meta, responseHeaders: { ...(result as { _meta?: { responseHeaders?: Record<string, unknown> } })._meta?.responseHeaders, 'x-proxied-replica': proxiedReplica, }, }, } as CallToolResult; } return result; } /** * Applies post-processing to a Gradio tool result: * - Image content filtering (conditionally) * - OpenAI-specific structured content * * This should be called after any custom transformations (like _mcpui handling) * to ensure consistent behavior across all Gradio tools. */ export function applyResultPostProcessing( result: CallToolResult, options: GradioToolCallOptions ): CallToolResult { // Strip image content if requested const filteredResult = stripImageContentFromResult(result, { enabled: !!options.stripImageContent, toolName: options.toolName, outwardFacingName: options.outwardFacingName, }); // For OpenAI MCP client, check if result contains a URL and set structuredContent if (options.sessionInfo?.clientInfo?.name === 'openai-mcp') { const extractedUrl = extractUrlFromContent(filteredResult.content); if (extractedUrl) { logger.debug({ tool: options.toolName, url: extractedUrl }, 'Setting structuredContent with extracted URL'); (filteredResult as CallToolResult & { structuredContent?: { url: string; spaceName?: string }; }).structuredContent = { url: extractedUrl, spaceName: options.spaceName, }; } } return filteredResult; }