Google Search Console MCP Server

An MCP server for Google Search Console that lets you ask Claude questions about your search data and get real answers. Not raw API rows. Actual analysis.

29 tools. OAuth or service account. Free and open source. Runs on your machine: your data goes straight from this computer to Google, and nothing passes through anyone else’s servers.

https://github.com/user-attachments/assets/ef131907-4461-4c6a-8080-770f8162dbf2

Also on YouTube.

Full setup guide with screenshots: suganthan.com/blog/google-search-console-mcp-server/

v2.5.0 update (August 2026): new tool image_page_audit closes the image SEO loop. The v2.3 suite tells you which pages fail in image search; this one fetches those pages from your own site and tells you why: alt text, filenames, dimension attributes, lazy loading on the LCP image, formats and weights, the ~250x200 indexing minimum, ImageObject/licensable schema, max-image-preview, and the metadata inside the image files (camera EXIF to strip, IPTC to keep, DigitalSourceType on AI images). It only ever fetches the URLs you give it. Launch post with the whole 8-tool workflow on real client data: “One Page Earned 102,657 Image Impressions and 2 Clicks”.

v2.4.0 update (August 2026): new tool genai_conversation_queries finds the AI conversations leaking into your query report. People reply to Google’s AI with things like “yes, go on”, Google logs every follow-up as a new query, and this tool sorts all of it into seven classified buckets with landing pages and a monthly timeline. Full method and findings: “Yes, Go On”: The AI Conversations Leaking Into Your Search Console.

See it in action

“Audit the images on these pages”

“Which of my queries are actually AI conversations?”

“How is my site doing?”

“What are my quick win keywords?”

“Which pages are cannibalising each other?”

“What content is slowly dying?”

“Which pages lost traffic and why?”

“How does my CTR compare to benchmarks?”

“How is my blog cluster performing?”

What you can ask

"What are my quick win keywords?"
"Which pages lost traffic this month and why?"
"What content is decaying?"
"Which pages are cannibalising each other?"
"Check for any SEO alerts in the last 7 days"
"Give me content recommendations"
"How does my CTR compare to benchmarks?"
"How is my /blog/ cluster performing?"
"Show me US mobile traffic for the last 90 days"
"Is /blog/my-post/ indexed? If not, why?"
"Generate a full performance report and save it"
"Show me a dashboard across all my sites"
"Submit this URL for indexing: https://mysite.com/new-post/"
"Batch submit all my new blog posts for indexing"
"List my sitemaps and their status"
"Verify that claim about my homepage clicks"

Quick start

One command setup (new in v2.3)

npx -y suganthan-gsc-mcp setup

The wizard signs you in with Google, verifies the connection with a live API call, lets you pick your property from a list, and writes the config for Claude Desktop and Claude Code. No config files to edit.

Read only by default: the standard consent screen asks for a single view permission. Choose full access during setup if you want the sitemap and URL submission tools.

For now you still need your own Google OAuth client JSON one time (steps 1 to 3 under Manual OAuth below); the wizard takes it from there. Built in Google sign in, with no Google Cloud steps at all, ships the moment Google finishes verifying the shared client.

Useful flags: --client desktop|code|both|print, --scopes readonly|full, --site , --secrets , --reauth, --force, --dry-run, --help.

One click desktop install

Prefer no terminal at all? Download the .mcpb bundle from the releases page and double click it. Claude Desktop installs the server with a small settings screen.

Option A: OAuth (manual)

1. Create a Google Cloud project and enable the Search Console API
2. Go to Credentials > Create Credentials > OAuth client ID, choose Desktop app
3. Download the client secrets JSON
4. Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "gsc": {
      "command": "npx",
      "args": ["-y", "suganthan-gsc-mcp"],
      "env": {
        "GSC_AUTH_MODE": "oauth",
        "GSC_OAUTH_SECRETS_FILE": "/path/to/client_secrets.json",
        "GSC_SITE_URL": "sc-domain:yoursite.com",
        "GSC_SCOPES": "readonly"
      }
    }
  }
}

First use opens a browser for Google sign in. Token is cached after that (locally, at ~/.gsc-mcp/). Set GSC_SCOPES to full if you want the submission tools; omit it and you get full access, matching pre 2.3 behaviour. Running from a git checkout instead of npm? Use "command": "node", "args": ["/path/to/Suganthans-GSC-MCP/dist/index.js"].

Option B: Service Account

1. Create a Google Cloud project and enable the Search Console API
2. Go to IAM & Admin > Service Accounts, create one, download the JSON key
3. Add the service account email to your GSC property (Settings > Users and permissions > Full access)
4. Add to your Claude Desktop config:

{
  "mcpServers": {
    "gsc": {
      "command": "node",
      "args": ["/path/to/Suganthans-GSC-MCP/dist/index.js"],
      "env": {
        "GSC_KEY_FILE": "/path/to/service-account.json",
        "GSC_SITE_URL": "sc-domain:yoursite.com"
      }
    }
  }
}

Generative AI (v2.4)

Google’s Generative AI performance report has no API, no BigQuery export, and no searchAppearance value. But Google counts every AI Mode follow-up as a brand-new query and folds AI Mode and AI Overviews into the web search type, so AI-conversation exhaust leaks into the regular query dimension with real impressions, positions and clicks. This tool mines it.

Indexing API (optional)

To use submit_url, submit_batch, and submit_sitemap:

1. Enable the Web Search Indexing API in your Google Cloud console
2. Your service account (or OAuth credentials) need owner-level access in Search Console

Note: Google officially says the Indexing API is for JobPosting and BroadcastEvent schema types. In practice, it processes requests for all page types.

Multi-site

For multiple properties, add GSC_SITE_URLS:

"env": {
  "GSC_SITE_URL": "sc-domain:primarysite.com",
  "GSC_SITE_URLS": "sc-domain:primarysite.com,sc-domain:secondsite.com"
}

You can also point a single query at any property your credentials can see without touching the config: advanced_search_analytics, genai_conversation_queries, and all 7 image-search analysis tools take an optional site_url parameter. “Which pages get image impressions but no clicks on sc-domain:secondsite.com?” just works.

All 29 tools

Analysis

Image SEO (v2.3 + v2.5)

These tools pass type=image to the GSC Search Analytics API, which most third-party tools never expose. They cover the visual-search surface end-to-end.

Indexing

Safety

What makes this different from other Google Search Console MCP servers

Analysis, not just API access. Most Google Search Console MCP servers wrap the raw API. This one ships with pre-built analysis: opportunity scoring, cannibalisation detection, decay tracking, CTR benchmarking, traffic drop diagnosis. You ask a question, it runs the analysis and tells you what to do.

Local and private. No hosted middleman, no account, no plan. The server runs on your machine, tokens are cached on your machine, and your Search Console data travels directly between your machine and Google. The developer operates no servers and receives nothing. Read only scope by default.

Hallucination guardrails. Every tool instructs Claude to base analysis only on returned data. Provenance metadata in every response. The verify_claim tool lets Claude fact-check its own numbers. Credit to Krinal Mehta for pushing this.

Visual dashboards. Results render as rich, interactive visualisations in Claude Desktop. Summary cards, colour coded indicators, bar charts, and tabbed sections. Not plain text dumps.

Fresh data. Uses dataState: 'all' so data matches the GSC dashboard, not 2-3 days stale.

Proactive, not reactive. Alerting, content recommendations, and scheduled reports catch problems before you think to look.

Environment variables

Full guide

Step-by-step setup with screenshots, use cases, and examples:

suganthan.com/blog/google-search-console-mcp-server/

Changelog

v2.5.1 Three fixes from the issue queue. The sitemaps tools now call sc-domain: properties as-is, which the Sitemaps API supports and always has; the old URL-prefix rewrite manufactured a 403, blamed the API for it, and on some accounts silently returned a stale URL-prefix property’s sitemap list instead of the domain property’s. The rewrite survives only as a fallback for accounts whose permission genuinely lives on a URL-prefix property. Credit to pauljlange for proving the premise wrong with raw API calls. Service-account mode now requests the same scope set as the OAuth flow, so submit_url and submit_batch work with service accounts and the GSC_SCOPES=readonly tier applies there too; the OAuth half of this shipped earlier via tomschmidty12-sys’s PR, after stuli1989 reported it and sent the first fix for both modes in #3. And the Google client libraries are bumped (googleapis 144 to 176, google-auth-library 10, gaxios 7) so token fetches use Node’s native fetch; the old node-fetch path failed to decompress Google’s gzip token responses on Node 22 and newer, which broke service accounts, most visibly on Windows. Credit to wilotas for the full diagnosis and the verified fix. Also fixed: the one-click desktop bundle. Every previous .mcpb shipped without the googleapis package internals (the packer silently pruned them), so the extension could never start, and “Unable to connect to extension server” was this bug, not your config. The v2.5.1 bundle carries the complete production dependency tree and was runtime-tested against the live API before upload, which is why it is bigger. Credit to VeloWulf for the report.

v2.5.0 Image page audit. image_page_audit fetches up to 5 pages from your own site and audits every image on them against the on-page factors that decide image-search performance: missing, empty, generic, filename-as-alt or duplicate alt text, camera-default and stock-agency-default filenames (shutterstock_524347192.jpg tells Google nothing), missing width/height attributes, lazy loading on the LCP candidate, srcset coverage, file format and weight (flags photos shipped as PNG and anything over 500KB), intrinsic dimensions against Google’s ~250x200 indexing minimum, ImageObject and licensable-field schema, primaryImageOfPage, max-image-preview, inline background images, and the metadata inside the image files via exifr: camera EXIF and GPS that should be stripped, IPTC Creator/Copyright/Caption that should survive your CMS, and XMP DigitalSourceType on AI-generated images. Returns per-image findings, page-level checks, and an ordered top_fixes list. Pairs with image_impressions_no_clicks and image_search_quick_wins: those name the pages, this names the reasons. Fetches only the URLs it is given, so the privacy model is unchanged: your data goes to Google and your own site and nowhere else. New dependencies: node-html-parser, image-size, exifr (all pure JS, no native builds). The 7 image-search analysis tools also gain the optional site_url override that advanced_search_analytics already had, so any of them can be pointed at any property your credentials can see without editing the config. Launch post with the full workflow on real client data: “One Page Earned 102,657 Image Impressions and 2 Clicks”.

v2.4.0 Generative AI conversation queries. genai_conversation_queries finds the AI conversation fragments hiding in your regular query data and sorts them into seven kinds: reply artefacts (“yes”, “go on”), pivot follow-ups (“what about resend?”), conversational questions, tracker probes, agent harnesses, pasted strings, and a review pile. Google counts every AI Mode follow-up as a brand new query, so these rows carry real impressions, positions and clicks, and the dedicated Generative AI report has no query view, which makes this the only query-level AI evidence available anywhere. One call classifies sixteen months of your queries, attaches landing pages via query and page grouping, and returns a monthly reply-artefact timeline. Plain Search Analytics API, no BigQuery, no new permissions. Full method and findings: the launch post. Sparked by Anastasia Kourou surfacing the queries with John Mueller confirming the mechanism, and by Ross Tavendale asking how to reverse engineer it.

v2.3.0 Image SEO suite and one command setup. 7 new tools that pass type=image to the GSC Search Analytics API, plus a type parameter on advanced_search_analytics covering all 6 GSC search surfaces (web, image, video, news, discover, googleNews). The image-search surface was invisible to most third-party SEO tools because they default to type=web and never expose the others; v2.3 makes it queryable end-to-end. Also new: npx suganthan-gsc-mcp setup, a wizard that signs you in, verifies the connection with a live call, and writes your Claude Desktop and Claude Code configs; a read only scope tier (GSC_SCOPES=readonly, now the setup default) so the standard consent asks for one view permission; and a one click Claude Desktop bundle (.mcpb) on the releases page.

v2.2.2 Published to npm as suganthan-gsc-mcp. Config can now use npx instead of a local checkout path.

v2.2.1 Fixed OAuth EADDRINUSE crash when multiple tool calls triggered concurrent authentication flows. The server now reuses the active auth session instead of spawning duplicate listeners. Thanks to Rushabh Rathod for finding and reporting this.

v2.2.0 Visual dashboard rendering. All analysis tools now produce rich, interactive visualisations in Claude Desktop with summary cards, colour coded indicators, bar charts, and tabbed sections instead of plain text output. No reinstall needed, just restart Claude Desktop.

v2.1.0 Added Indexing API tools: submit_url, submit_batch, submit_sitemap, list_sitemaps. Request Google to crawl and index pages directly from Claude.

v2.0.0 Added OAuth authentication, advanced search analytics, check_alerts, content_recommendations, generate_report, multi_site_dashboard, verify_claim. Server grew from 10 to 16 tools.

v1.1.0 Added hallucination guardrails: explicit prompts in tool descriptions, data provenance metadata in responses, and verify_claim self-checking tool. Thanks to Krinal Mehta for the feedback.

v1.0.0 Initial release with 10 analysis tools and service account authentication.

Licence

Apache 2.0. See LICENSE and NOTICE for details. Use it, fork it, build on it. Just keep the attribution.

---

Built by Suganthan Mohanadasan

I write about organic search, AI, and the tools I build on suganthan.com.

Posts about this server

Outgrowing the Search Console API?

The BigQuery MCP Server reads your Search Console bulk export instead of the API, so large sites keep the rows the API cuts and you can see how much of your traffic sits behind anonymised queries. It has 33 tools, including ML forecasting, anomaly detection, and GA4 revenue per keyword. Setup guide.

Follow along

If you find it useful, star it.