GSC-MCP

Remote MCP server that makes Google Search Console data available in Claude, ChatGPT, and Cursor — based on Google's full data export instead of the limited API.

Domain: gsc2mcp.drossmedia.de · Hosting: netcup root server, Nuremberg, with Cloudflare in front

Status: Implementation started. The network-independent core logic is available as tested packages (core, analytics, db, gsc-client) along with the scaffolding for the MCP server (apps/app) and sync worker (apps/worker). The network-side wiring — MCP transport, OAuth provider, Google integration — follows with Phase 0/1 of the roadmap once the domain and GCP OAuth client are ready.

What It's About

The Search Console API is tightly limited: 16 months of history, 25,000 rows per request, roughly 50,000 rows per day per search type. Every single tool on the market uses this API — even those that store data afterward. They are storing a sample.

Google's Bulk Data Export has none of these limits: complete search data, daily, no API quota, unlimited retention. What it lacks is everything else — no interface, no analysis, no agent access.

GSC-MCP closes exactly this gap and adds a deterministic analysis engine: change attribution, seasonally adjusted anomaly detection, site-specific CTR curves, cannibalization time series. The numbers are reproducible, not estimated.

How big the difference is has been measured: For sc-domain:aip.aero, the top 100 search queries account for 8.3% of clicks — roughly 92% of activity lies outside what a tool with a row cap can show.

Related MCP server: searchconsole-mcp

Documentation

Architecture in Brief

Claude / ChatGPT / Cursor
        │  Streamable HTTP + OAuth 2.1
        ▼
   Cloudflare (TLS, DDoS, WAF, Rate Limiting)
        │
        ▼
   netcup Root Server · Nürnberg
     Caddy → app (MCP + OAuth AS) · web · worker
     PostgreSQL 17 (Warehouse, pg-boss, Rate-Budget)
        │
        ├──▶ Search Console API      (einmaliger 16-Monats-Backfill)
        └──▶ BigQuery Bulk Export    (laufend, vollständig, ohne API-Quote)

A second hostname gsc2mcp-direct.drossmedia.de runs without Cloudflare directly on the server — for customers whose procurement does not allow a US data processor, and as a fallback path in case of a proxy outage.

Project Structure

Monorepo, npm workspaces, TypeScript strict, Vitest.

The canonical DDL is in packages/db/migrations/0001_init.sql and is validated against real PostgreSQL; docs/03 explains it.

Development

npm install
npm run typecheck   # tsc --build über alle Packages
npm test            # Vitest über packages/*/test und apps/*/test

CI (GitHub Actions) additionally checks the migration against a real PostgreSQL, the concept documents (links, consistency), and the overview page in both themes. See docs/10-repo-struktur.md and CLAUDE.md.

Next Steps

The critical path is not the code, but the Google OAuth verification for the sensitive scope webmasters.readonly — it typically takes several weeks and requires a domain and privacy policy. Details in docs/02-auth.md and docs/09-roadmap.md.

Before Phase 2, it must also be decided where the Bulk Data Export delivers — to the customer's BigQuery project or to ours. The trade-off is discussed in docs/12-wettbewerb-usp.md.