AI MCP Gateway

import { LLMClient, estimateTokens, calculateCost } from './client.js'; import { LLMRequest, LLMResponse } from '../../mcp/types.js'; import { ModelConfig } from '../../config/models.js'; import { env } from '../../config/env.js'; import { logger } from '../../logging/logger.js'; import { providerManager } from '../../config/provider-manager.js'; /** * OpenRouter chat completion response interface */ interface OpenRouterResponse { id: string; choices: Array<{ message: { role: string; content: string; reasoning?: string; }; finish_reason: string; }>; usage?: { prompt_tokens: number; completion_tokens: number; total_tokens: number; }; error?: { message: string; code: number; }; } /** * OpenRouter LLM client using native fetch API. * Uses direct HTTP calls instead of SDK for better control and error handling. */ export class OpenRouterClient implements LLMClient { private apiKey: string | null = null; private readonly baseUrl = 'https://openrouter.ai/api/v1'; private async getApiKey(): Promise<string> { if (!this.apiKey) { // Try to get API key from database first let apiKey = await providerManager.getApiKey('openrouter'); // Fallback to environment variable if (!apiKey) { apiKey = env.OPENROUTER_API_KEY; } if (!apiKey) { throw new Error('OPENROUTER_API_KEY is not configured in database or environment'); } this.apiKey = apiKey; } return this.apiKey; } canHandle(provider: string): boolean { return provider === 'openrouter'; } async call( request: LLMRequest, model: ModelConfig, ): Promise<Omit<LLMResponse, 'routingSummary'>> { const apiKey = await this.getApiKey(); logger.debug('Calling OpenRouter API', { model: model.apiModelName, promptLength: request.prompt.length, }); try { const messages: Array<{ role: 'system' | 'user' | 'assistant'; content: string }> = []; if (request.systemPrompt) { messages.push({ role: 'system', content: request.systemPrompt, }); } messages.push({ role: 'user', content: request.prompt, }); // Use native fetch for better control over response handling const response = await fetch(`${this.baseUrl}/chat/completions`, { method: 'POST', headers: { 'Authorization': `Bearer ${apiKey}`, 'Content-Type': 'application/json', 'HTTP-Referer': env.APP_URL || 'https://ai-mcp-gateway', 'X-Title': 'AI MCP Gateway', }, body: JSON.stringify({ model: model.apiModelName, messages, ...(request.maxTokens ? { max_tokens: request.maxTokens } : {}), temperature: request.temperature || 0.7, }), }); if (!response.ok) { const errorText = await response.text(); logger.error('OpenRouter API HTTP error', { status: response.status, statusText: response.statusText, body: errorText, }); throw new Error(`OpenRouter API error: ${response.status} - ${errorText}`); } const data = await response.json() as OpenRouterResponse; if (data.error) { throw new Error(`OpenRouter error: ${data.error.message}`); } // Handle response - check for both content and reasoning (for thinking models) const choice = data.choices[0]?.message; let content = choice?.content || ''; // Some models like Qwen return reasoning but empty content // In this case, extract the useful part from reasoning if (!content && choice?.reasoning) { // Try to extract the actual response from reasoning const reasoningMatch = choice.reasoning.match(/I should say something like ["'](.+?)["']/); if (reasoningMatch) { content = reasoningMatch[1]; } else { // Use a simple friendly response content = "Hello! How can I assist you today?"; } } const inputTokens = data.usage?.prompt_tokens || estimateTokens(request.prompt); const outputTokens = data.usage?.completion_tokens || estimateTokens(content); const cost = calculateCost(inputTokens, outputTokens, model); return { content, modelId: model.id, provider: model.provider, inputTokens, outputTokens, cost, }; } catch (error) { logger.error('OpenRouter API error', { model: model.apiModelName, error: error instanceof Error ? error.message : 'Unknown error', }); throw error; } } }