OpenRouter Agents MCP Server

require('dotenv').config(); const pkg = require('./package.json'); const config = { server: { // Support both SERVER_PORT and PORT, prefer SERVER_PORT if present port: process.env.SERVER_PORT || process.env.PORT || 3002, name: "openrouter_agents", version: pkg.version, // Add a key for basic server authentication (optional) apiKey: process.env.SERVER_API_KEY || null, requireHttps: process.env.REQUIRE_HTTPS === 'true', publicUrl: process.env.PUBLIC_URL || `${process.env.REQUIRE_HTTPS === 'true' ? 'https' : 'http'}://localhost:${process.env.SERVER_PORT || process.env.PORT || 3002}`, // Server startup behavior allowStartWithoutDb: process.env.ALLOW_START_WITHOUT_DB === 'true', startupTimeoutMs: parseInt(process.env.STARTUP_TIMEOUT_MS, 10) || 30000 }, openrouter: { apiKey: process.env.OPENROUTER_API_KEY, baseUrl: "https://openrouter.ai/api/v1" }, models: { // Allow overriding planning model; provide a generally-available safe default planning: process.env.PLANNING_MODEL || "google/gemini-2.5-pro", // Default planning/synthesis model planningCandidates: (process.env.PLANNING_CANDIDATES || "openai/gpt-5-chat,google/gemini-2.5-pro,anthropic/claude-sonnet-4") .split(',').map(s=>s.trim()).filter(Boolean), useDynamicCatalog: process.env.USE_DYNAMIC_CATALOG === 'true', // Define models with domain strengths // Accept either JSON array of objects or CSV of model ids in env vars highCost: process.env.HIGH_COST_MODELS ? (function parseHighCost(val){ try { const parsed = JSON.parse(val); if (Array.isArray(parsed)) return parsed; } catch(_) {} // CSV fallback -> wrap ids as objects without explicit domains return String(val).split(',').map(s=>s.trim()).filter(Boolean).map(name=>({ name, domains: ["general"] })); })(process.env.HIGH_COST_MODELS) : [ { name: "x-ai/grok-4", domains: ["reasoning", "technical", "general", "creative"] }, { name: "openai/gpt-5-chat", domains: ["reasoning", "technical", "general"] }, { name: "google/gemini-2.5-pro", domains: ["reasoning", "technical", "general"] }, { name: "anthropic/claude-sonnet-4", domains: ["reasoning", "technical", "general"] }, { name: "qwen/qwen3-coder", domains: ["coding", "editing", "technical"] }, { name: "qwen/qwen3-235b-a22b-2507", domains: ["general", "reasoning", "technical", "coding"] } ], lowCost: process.env.LOW_COST_MODELS ? (function parseLowCost(val){ try { const parsed = JSON.parse(val); if (Array.isArray(parsed)) return parsed; } catch(_) {} return String(val).split(',').map(s=>s.trim()).filter(Boolean).map(name=>({ name, domains: ["general"] })); })(process.env.LOW_COST_MODELS) : [ { name: "deepseek/deepseek-chat-v3.1", domains: ["general", "reasoning", "technical", "coding"] }, { name: "z-ai/glm-4.5v", domains: ["general", "multimodal", "vision", "reasoning"] }, { name: "z-ai/glm-4.5-air", domains: ["coding", "technical", "reasoning"] }, { name: "openai/gpt-oss-120b", domains: ["general", "reasoning", "search"] }, { name: "inception/mercury", domains: ["general", "creative", "technical"] }, { name: "baidu/ernie-4.5-vl-424b-a47b", domains: ["general", "creative"] }, { name: "google/gemini-2.5-flash", domains: ["coding", "editing", "technical"] } ], // Define a tier for potentially simpler tasks (adjust models as needed) veryLowCost: process.env.VERY_LOW_COST_MODELS ? (function parseVeryLow(val){ try { const parsed = JSON.parse(val); if (Array.isArray(parsed)) return parsed; } catch(_) {} return String(val).split(',').map(s=>s.trim()).filter(Boolean).map(name=>({ name, domains: ["general"] })); })(process.env.VERY_LOW_COST_MODELS) : [ { name: "openai/gpt-5-nano", domains: ["general", "reasoning", "creative"] } ], // Add a model specifically for classification tasks if needed, or reuse planning model classification: process.env.CLASSIFICATION_MODEL || "openai/gpt-5-mini", // Default ensemble size for research agent model ensembles ensembleSize: parseInt(process.env.ENSEMBLE_SIZE, 10) || 2, // Max research iterations (initial + refinements) maxResearchIterations: parseInt(process.env.MAX_RESEARCH_ITERATIONS, 10) || 2, // Default to 1 initial + 1 refinement // Parallelism for concurrent sub-queries parallelism: parseInt(process.env.PARALLELISM, 10) || 4, // Enforce a minimum max_tokens to avoid truncation across all calls minMaxTokens: parseInt(process.env.MIN_MAX_TOKENS, 10) || 2048, // Model-aware adaptive token limits (dynamically adjusted based on model capabilities) tokens: { synthesis: { min: parseInt(process.env.SYNTHESIS_MIN_TOKENS, 10) || 4000, fallbackMax: parseInt(process.env.SYNTHESIS_MAX_TOKENS, 10) || 16000, perSubQuery: parseInt(process.env.TOKENS_PER_SUBQUERY, 10) || 800, perDocument: parseInt(process.env.TOKENS_PER_DOC, 10) || 500 }, research: { min: parseInt(process.env.RESEARCH_MIN_TOKENS, 10) || 2000, fallbackMax: parseInt(process.env.RESEARCH_MAX_TOKENS, 10) || 8000 }, planning: { min: parseInt(process.env.PLANNING_MIN_TOKENS, 10) || 1000, fallbackMax: parseInt(process.env.PLANNING_MAX_TOKENS, 10) || 4000 } } }, // Database configuration for knowledge base using PGLite database: { dataDirectory: process.env.PGLITE_DATA_DIR || "./researchAgentDB", vectorDimension: 384, // Dimension for the embeddings from all-MiniLM-L6-v2 cacheTTL: parseInt(process.env.CACHE_TTL_SECONDS, 10) || 3600, // 1 hour in seconds // Enhanced PGLite configuration databaseUrl: process.env.PGLITE_DATABASE_URL || null, // Override auto-detected URL relaxedDurability: process.env.PGLITE_RELAXED_DURABILITY === 'false' ? false : true, maxRetryAttempts: parseInt(process.env.PGLITE_MAX_RETRY_ATTEMPTS, 10) || 3, retryDelayBaseMs: parseInt(process.env.PGLITE_RETRY_DELAY_BASE_MS, 10) || 200, // Initialization behavior initTimeoutMs: parseInt(process.env.PGLITE_INIT_TIMEOUT_MS, 10) || 30000, retryOnFailure: process.env.PGLITE_RETRY_ON_FAILURE === 'true', allowInMemoryFallback: process.env.PGLITE_ALLOW_IN_MEMORY_FALLBACK !== 'false' // Default true for backwards compat }, // Local indexing/search configuration (opt-in by default) indexer: { enabled: process.env.INDEXER_ENABLED === 'false' ? false : true, autoIndexReports: process.env.INDEXER_AUTO_INDEX_REPORTS === 'true', autoIndexFetchedContent: process.env.INDEXER_AUTO_INDEX_FETCHED === 'true', embedDocs: process.env.INDEXER_EMBED_DOCS !== 'false', maxDocLength: parseInt(process.env.INDEXER_MAX_DOC_LENGTH, 10) || 8000, bm25: { k1: Number(process.env.INDEXER_BM25_K1) || 1.2, b: Number(process.env.INDEXER_BM25_B) || 0.75 }, weights: { bm25: Number(process.env.INDEXER_WEIGHT_BM25) || 0.7, vector: Number(process.env.INDEXER_WEIGHT_VECTOR) || 0.3 }, stopwords: (process.env.INDEXER_STOPWORDS || '').split(',').map(s => s.trim()).filter(Boolean), rerankEnabled: process.env.INDEXER_RERANK_ENABLED === 'true', rerankModel: process.env.INDEXER_RERANK_MODEL || null }, // Configuration for where to save full research reports reportOutputPath: process.env.REPORT_OUTPUT_PATH || './research_outputs/' }; // MCP feature toggles (opt-in by default; can be disabled via env) config.mcp = { features: { prompts: process.env.MCP_ENABLE_PROMPTS === 'false' ? false : true, resources: process.env.MCP_ENABLE_RESOURCES === 'false' ? false : true } }; config.mcp.mode = (process.env.MODE || 'ALL').toUpperCase(); // Experimental modes config.modes = { hyper: process.env.HYPER_MODE === 'true' }; // MCP transport preferences config.mcp.transport = { streamableHttpEnabled: process.env.MCP_STREAMABLE_HTTP_ENABLED === 'false' ? false : true }; // Prompt strategy configuration config.prompts = { compact: process.env.PROMPTS_COMPACT === 'false' ? false : true, // default compact prompts on requireUrls: process.env.PROMPTS_REQUIRE_URLS === 'false' ? false : true, confidenceScoring: process.env.PROMPTS_CONFIDENCE === 'false' ? false : true }; // Simple tool aliasing (short params) for minimal token overhead config.simpleTools = { enabled: process.env.SIMPLE_TOOLS === 'false' ? false : true }; // Tool recursion/chaining configuration // Allows agent to execute multiple tools in a single call with depth limiting config.toolRecursion = { enabled: process.env.MAX_TOOL_DEPTH !== '0', maxDepth: parseInt(process.env.MAX_TOOL_DEPTH, 10) || 3 // Default: 3 levels, set to 0 to disable }; // Async job processing - optimized for parallel batch research config.jobs = { concurrency: parseInt(process.env.JOBS_CONCURRENCY, 10) || 4, // ↑ from 2 for better parallelism heartbeatMs: parseInt(process.env.JOB_HEARTBEAT_MS, 10) || 2000, // ↓ from 5000 for faster stale detection leaseTimeoutMs: parseInt(process.env.JOB_LEASE_TIMEOUT_MS, 10) || 30000, // ↓ from 60000 for faster recovery batchEventLimit: parseInt(process.env.JOB_BATCH_EVENT_LIMIT, 10) || 500, // SSE event limit per batch poll ssePollingMs: parseInt(process.env.JOB_SSE_POLLING_MS, 10) || 500 // ↓ from 1000 for lower latency }; // Structured logging configuration (MCP-compliant) config.logging = { // Log level filtering: 'debug' | 'info' | 'warn' | 'error' (default: info) level: (process.env.LOG_LEVEL || 'info').toLowerCase(), // Output mode: 'stderr' | 'mcp' | 'both' (default: stderr) // - stderr: Traditional stderr logging (compatible with all clients) // - mcp: Use MCP SDK sendLoggingMessage notifications (client can filter) // - both: Output to both channels output: process.env.LOG_OUTPUT || 'stderr', // JSON format for machine parsing (default: false for human-readable) json: process.env.LOG_JSON === 'true' }; // Advanced caching and cost optimization config.caching = { // Semantic result caching results: { enabled: process.env.RESULT_CACHING_ENABLED !== 'false', ttlSeconds: parseInt(process.env.RESULT_CACHE_TTL, 10) || 7200, // 2 hours maxEntries: parseInt(process.env.RESULT_CACHE_MAX_ENTRIES, 10) || 1000, similarityThreshold: parseFloat(process.env.CACHE_SIMILARITY_THRESHOLD) || 0.85 }, // Model response caching models: { enabled: process.env.MODEL_CACHING_ENABLED !== 'false', ttlSeconds: parseInt(process.env.MODEL_CACHE_TTL, 10) || 3600, // 1 hour maxEntries: parseInt(process.env.MODEL_CACHE_MAX_ENTRIES, 10) || 500 }, // Cost optimization strategies optimization: { preferredLowCostModels: ['deepseek/deepseek-chat-v3.1', 'qwen/qwen3-coder', 'z-ai/glm-4.5v'], visionModels: ['z-ai/glm-4.5v', 'google/gemini-2.5-flash', 'openai/gpt-5-nano'], codingModels: ['qwen/qwen3-coder', 'z-ai/glm-4.5-air', 'deepseek/deepseek-chat-v3.1'], complexReasoningModels: ['deepseek/deepseek-chat-v3.1', 'qwen/qwen3-235b-a22b-2507', 'nousresearch/deephermes-3-mistral-24b-preview'], costThresholds: { simple: 0.0000005, // Max cost per token for simple queries moderate: 0.000002, // Max cost per token for moderate queries complex: 0.000015 // Max cost per token for complex queries } } }; // Core abstractions (Convergence Plan v2.0) config.core = { // Enable new consolidated handlers (gradual migration) handlers: { enabled: process.env.CORE_HANDLERS_ENABLED === 'true', // Which domains use new handlers (others fall back to tools.js) domains: (process.env.CORE_HANDLER_DOMAINS || '').split(',').filter(Boolean) }, // Signal protocol configuration signal: { enabled: process.env.SIGNAL_PROTOCOL_ENABLED === 'true', maxHistorySize: parseInt(process.env.SIGNAL_MAX_HISTORY, 10) || 1000 }, // RoleShift bidirectional protocol roleShift: { enabled: process.env.ROLESHIFT_ENABLED === 'true', timeout: parseInt(process.env.ROLESHIFT_TIMEOUT_MS, 10) || 60000 }, // Schema registry options schemas: { strictValidation: process.env.STRICT_SCHEMA_VALIDATION === 'true' } }; // Version information config.version = pkg.version; config.mcpSpec = { stable: '2025-06-18', draft: '2025-11-25', features: ['SEP-1686', 'SEP-1577', 'SEP-1036', 'SEP-1865'] }; module.exports = config;