System Architecture Overview
Key Subsystems
1. Fluent Domain Architecture (src/tools/fluent/)
Search Console MCP v2.0 groups all operations into 7 clean domain modules:
sites.ts: Site property management & multi-account configuration (sites_list,sites_manage,accounts_manage).sitemaps.ts: Sitemap listing, submission, and removal (sitemaps_list,sitemaps_submit,sitemaps_delete).analytics.ts: Unified search queries, PoP comparison, trends, anomalies, and drop attribution (analytics_query,analytics_compare,analytics_anomalies).inspection.ts: URL inspection & Core Web Vitals audit (inspection_inspect,pagespeed_analyze).indexing.ts: Instant URL indexing via IndexNow or Google/Bing APIs (indexing_submit,indexing_status).seo.ts: Automated SEO audits and keyword research (seo_audit,seo_keywords_research,schema_validate).health.ts: Cross-platform health checks & engine comparison (site_health_check,compare_engines).
2. Parallel Multi-Engine Runner (src/common/utils/parallel.ts)
Multi-engine queries (engine: "all") run concurrently using Promise.allSettled. If one engine encounters a temporary auth issue or timeout, the remaining engines return valid data without failing the overall query. This architecture reduces multi-engine latency by 50%+.
3. Fallback Router Layer (src/legacy/fallback-router.ts)
To ensure 100% backward compatibility for all ~96 legacy tool names (bing_sites_list, seo_quick_wins, sitemaps_get, bing_index_now, etc.), the Fallback Router intercepts legacy invocations and delegates them directly to the corresponding Fluent handler.
Read full Backward Compatibility Documentation →
Security & Privacy Design
- Local Execution: The server runs locally on your machine. Your API keys and search data are never sent to third-party tracking servers.
- Encrypted Token Vault: Credentials are stored in system Keychains (macOS Keychain, Windows Credential Manager) with AES-256-GCM hardware-bound file encryption fallback.
- Deterministic Server-Side Calculations: SEO math (cannibalization scores, standard deviations, moving averages) is calculated deterministically on the server before passing cleaned results to the LLM context window.