The short answer
Google does not publish a Search Console MCP server, so most options are community projects you run yourself with a Google Cloud key. DuckDive is a hosted MCP server instead: you sign in with Google, link each site's Search Console property, and your agent reads each site's clicks, impressions, CTR and average position, plus its leading queries and pages.
- No Cloud project, no JSON key file, no service account email to add property by property.
- Your agent asks one site at a time and loops across all of them; the longest range it can request is 12 months.
- It cannot split queries by page, country or device. For that, the raw Search Console API is still the tool.
Google has no official Search Console MCP server
Search for one and you land on GitHub, for good reason. A fact-check of official marketing MCP servers published by Digital Applied on November 2, 2026 found no Google-published server for Search Console, only community projects, as of November 5, 2026. Google's Analytics team is a different story, and we compare its GA server with a hosted connector on a separate page.
The community servers are real work by real people, and two of them show the two common shapes well:
- ahonn/mcp-server-gsc is MIT licensed, runs through
npxon Node 18 or later, and authenticates with a service account JSON key that its config points to throughGOOGLE_APPLICATION_CREDENTIALS. Its one tool,search_analytics, takes any Search Console dimension and returns up to 25,000 rows. - AminForou/mcp-gsc is MIT licensed, runs in Python through
uvx, and offers either a desktop OAuth client or a service account. Both routes need a Google Cloud project with the Search Console API turned on. It adds URL inspection and sitemap tools.
For one site and a taste for running things locally, either is fine. The friction starts with many properties, a teammate who wants the same access, or a phone chat app that cannot start a process on your laptop.
That is the problem DuckDive is built around. You connect once with your Google login, DuckDive imports each linked property's history and keeps it, and your agent reaches it through one remote address, https://api.duckdive.ai/mcp. Search Console sits beside the site's other sources, such as GA4 sessions and conversions, so one conversation covers the search and what happened after the click.
The 16-month window and what happens to older data
Search Console only reaches back so far. SEOTesting's guide to Search Console data limitations puts it plainly: the data "goes back 16 months", and older days are deleted. The API has served the same 16 months as the web interface since 2018, as Search Engine Journal reported when Google extended it.
Put that on a calendar. On November 16, 2026, the oldest day Google still holds is in July 2025. This November against last November works. Against November 2024 it does not: those days are gone, and no tool, hosted or local, can import them now.
When you link a property, DuckDive pulls in the history Google still serves and stores it. Through your agent, the longest range the Search Console tools accept today is the last 12 months, so plan your questions around a year.
If you need an archive that outlives Google's window, Google's own route is the bulk data export to BigQuery, which schedules a daily export of your performance data. Set it up for any site whose seasons you want to compare over years.
Ask your agentShow monthly clicks for Thistledown Bridal over the last 12 months and name its seasonal peak.
Tools it calls search get_search_console_summary
A bridal shop tests this limit well. Its demand is seasonal, so one year holds exactly one peak. The summary arrives as weekly points, each week starting on a Monday, so the agent adds up the weekly points by month (a week that spans two months lands in one of them, which makes each monthly figure approximate). That is close enough to name the busiest month. If you need precision, ask for the three peak weeks instead. Comparing this year's peak with last year's needs data older than your agent can reach.
Hosted sign-in versus a service account key
Per the two READMEs above, the key route goes: pick a Google Cloud project, turn on the Search Console API, create a service account, download its JSON key. Then add the service account's email to every property, as a "property administrator" in one README and with full access in the other. Repeat for each new site.
Think hard about the key itself. Google Cloud's guidance on managing service account keys says keys "don't have an expiry time and stay valid until you delete them", and that a leaked key lets someone "gain a foothold in your environment." Its advice is to pick a safer way to authenticate whenever you can.
The hosted route swaps all of that for one Google screen. Your AI app only ever holds a DuckDive token issued over OAuth; nothing in your config file grants Google access on its own. DuckDive sees the properties your Google login already sees. Changed your mind? Revoking the connection under AI Connectors in your DuckDive account, or removing DuckDive from the AI app, stops every later request straight away.
| Question | DuckDive, hosted | Community server with a key |
|---|---|---|
| Google Cloud project | Not needed, Google sign-in only | Needed, with the Search Console API on |
| Credential you look after | None, your AI app holds a token | A JSON key that never expires by default |
| Access per property | Uses your login's existing access | Add the service account email to each property |
| Reachable from web and phone chat | Yes, one remote address | Partly, local unless you host it |
| Query by page, country, device | No | Yes, every API dimension; better fit here |
| Rows per answer | Up to 100 | Up to 25,000 in ahonn's server |
| GA4, Ads and LSA on the same site | Yes | No, Search Console only |
| Cost | Free for one site, then per site | Free, MIT licensed code |
Sources: ahonn/mcp-server-gsc README and AminForou/mcp-gsc README, checked November 16, 2026.
One more difference shows up past a handful of sites. A local server calls Google live for every question, against Search Console API usage limits counted per site, per user and per project. DuckDive answers from its stored copy, so thirty sites do not mean thirty live fetches. More on that in stored history versus live API servers.
What the three Search Console tools return
Three Search Console tools, plus the site lookup your agent runs first. Each data tool takes one website id and one preset window, the shortest a week back and the longest a year back (7, 14, 30 or 90 days, or 6 or 12 months).
list_websites- Every website your account can see, with the data sources connected to each. The agent uses it to get the id for each site before asking anything else.
get_search_console_summary- Totals for the range plus a series of weekly points, each dated to the Monday that starts the week, with clicks, impressions, CTR and average position in every point. This is the tool for trends, dips and seasonality. There are no daily points, even over 14 days.
get_search_console_keywords- The site's top Google search queries, ranked by clicks, up to 100 rows. Every row carries its own clicks, impressions, CTR and average position.
get_search_console_pages- The site's top URLs in Google results, ranked by clicks, up to 100 rows, with the same four numbers on each row.
The metrics mean what Google says they mean in its Performance report help. A click is a visit from a Google result. An impression is your site appearing in a result. CTR is clicks divided by impressions. Average position is "the average position of the topmost result from your site", which is why a site with two results on one query counts only its higher one.
Ranked by clicks
Both list tools sort by clicks. A query at position 11 with zero clicks never makes a top-100 list, so page-two hunting finds only queries that already pull some traffic.
Worked example: queries hovering just off page one across five sites
Positions 8 to 12 are the classic shortlist for a title rewrite, a stronger intro or a few internal links: the page already ranks and only needs a nudge. In Search Console that means five properties, five date pickers and five sorted tables. Through your agent it is one request.
Ask your agentAcross all my sites, list queries in each site's top 100 that averaged position 8 to 12 over the last 30 days.
Tools it calls list_websites get_search_console_keywords
Across all my sites, list queries in each site's top 100 that averaged position 8 to 12 over the last 30 days.
list_websitesfinds five websites, all with Search Console linkedget_search_console_keywordstop 100 queries by clicks for each site, last 30 days (five calls)
Six queries across four of your five sites sit between positions 8 and 12. Thistledown Bridal's dress-ordering query is closest to page one and already earns the most clicks in the group, so start there.
| Site and query | Avg position | Clicks |
|---|---|---|
| Thistledown Bridal: how far ahead to order a wedding dress | 8.2 | 96 |
| Brightfern Garden Supply: when to plant garlic in fall | 8.6 | 74 |
| Larchgate Tile Works: heated bathroom floor cost | 9.1 | 52 |
| Oxbow Canoe Rentals: overnight canoe trip checklist | 10.3 | 18 |
| Brightfern Garden Supply: raised bed soil mix ratio | 11.4 | 29 |
| Thistledown Bridal: bridal alterations timeline | 11.9 | 21 |
Quincewood Pottery Studio has none: its queries rank either 1 to 4 or below 20. This list only covers queries that earned clicks, so queries in this band with no clicks at all are not in it.
Notice the last line: a good agent says what the data leaves out. For a publisher with a stable of content sites, the same question is the start of deciding where next month's writing goes; our page for content site portfolios takes that further. Agencies that run this across a client roster after an algorithm update should read the core update triage page for SEO agencies.
Build your own page-two shortlist from real numbers
Link your Search Console properties and ask the same question about every site you run.
Free for 2 sites · No credit card
Questions you cannot ask of this data yet
Better to hear this before you sign up. DuckDive's tools do not answer these today:
- Which page ranks for which query. Queries and pages come back as two separate lists. There is no query-by-page table. Google's own Search Analytics query method can group by both, though its usage limits page calls a page and query grouping the most expensive request you can make.
- Country and device splits. Not exposed. Use the Performance report or a server that passes every API dimension through.
- Custom dates. No "since the update on the 3rd". Pick the nearest fixed range.
- More than 100 rows. Each list stops at 100. A large site's long tail stays in Search Console.
- Every query Google saw. Search Console itself does not report all queries; SEOTesting's limitations guide notes this mostly hits long-tail searches. No connector can return what Google withholds.
- URL inspection, sitemaps and indexing. DuckDive has no tools for them. The AminForou server above does, if your work is technical SEO.
You can still get close on the first one. Ask for both lists and let the agent pair them, as long as it labels the pairing as a guess:
Ask your agentFor Oxbow Canoe Rentals, put the top 20 pages and the top 20 queries by search clicks over the last 6 months side by side, then guess which queries land on which page and mark every match as a guess.
Tools it calls search get_search_console_pages get_search_console_keywords
So who should skip DuckDive here? Someone with one site who wants every row and dimension and is happy to manage a key, or someone whose daily job is indexing and sitemaps. A community server or the API fits them better, and it can run beside DuckDive in the same agent. Our roundup of marketing data MCP servers lays out the wider field.
Worked example: impressions up, click-through rate down
This pattern sets off alarms, often for no reason. CTR is a ratio, so it falls whenever impressions grow faster than clicks. The real question is whether you lost clicks or gained exposure lower down.
Ask your agentFor Brightfern Garden Supply, show clicks, impressions and CTR week by week for the last 90 days and point out when CTR fell while impressions rose.
Tools it calls search get_search_console_summary get_search_console_keywords
For Brightfern Garden Supply, show clicks, impressions and CTR week by week for the last 90 days and point out when CTR fell while impressions rose.
searchfinds the Brightfern Garden Supply websiteget_search_console_summaryclicks, impressions, CTR and position by week, 13 weekly points from the week of Aug 17 to the week of Nov 9get_search_console_keywordsleading queries over the past 30 daysget_search_console_keywordsleading queries over the past 90 days, to spot newcomers
CTR slid in each of the three stretches below while impressions climbed. I grouped the 13 weekly points into runs of four, four and five weeks and show the average week for each, so the longer last run does not inflate its totals. Impressions per week rose from 21,000 to 24,375 to 29,500 across the three runs. Clicks per week barely moved, down about 5% from the first run to the last. Average position got worse at the same time, which says Google is showing the site for more searches, lower on the page.
| Weeks of | Clicks a week | CTR | Avg pos. |
|---|---|---|---|
| Aug 17 to Sep 7 | 651 | 3.1% | 14.2 |
| Sep 14 to Oct 5 | 658 | 2.7% | 16.0 |
| Oct 12 to Nov 9 | 620 | 2.1% | 18.9 |
Four queries in the last-30-day list are missing from the 90-day list, all autumn how-to questions such as overwintering dahlias, all at position 17 or worse. This is a wider net, not a lost ranking. Nothing needs fixing unless clicks start falling too.
Use the same three numbers to read any version of this pattern:
- Impressions up, CTR down, position worse, clicks steady. More exposure at lower spots. Usually healthy, often seasonal.
- Impressions steady, CTR down, position steady. The results page changed, or your snippet stopped earning the click. Check titles on the top pages.
- Clicks down along with position. That is a real ranking loss. Take it to the multi-site traffic drop walkthrough, which starts by asking whether one site, some or all of them fell.
One caution: Google's help says the newest data "can be preliminary" and may still change, so a dip in the latest week can partly fill back in. The week of Nov 9 above is the one to recheck before acting on it.
Domain and URL-prefix properties
The property you link decides what your agent sees. Google's guide to adding a Search Console property describes two kinds:
- Domain property. Covers the domain with every subdomain, both http and https, and every path. It has to be verified with a DNS record. In the API its name looks like
sc-domain:example.com. - URL-prefix property. Covers only URLs that start with the exact prefix, protocol included. A separate property is needed for each subdomain or protocol. In the API its name looks like
https://www.example.com/.
The trap: link a www prefix property and a blog on blog. or a shop on shop. drops out of every answer. Link the domain property when one exists, and a prefix property only on purpose, such as a client who owns one folder of a larger site.
In DuckDive, each website links one Search Console site, and no Search Console site can feed two DuckDive websites at once. Try to attach one elsewhere and the agent asks you before moving it.
Ask your agentList the Search Console sites my Google login can reach, mark which ones are domain properties and which are URL-prefix, and tell me which DuckDive website each is linked to, if any.
Tools it calls list_available_sources list_websites
The answer is reliable because list_available_sources returns each Search Console site under the same name the API uses, so the sc-domain: prefix gives a domain property away at a glance.
Before you link your Search Console sites
Setup is one sign-in and one prompt. Check these first to avoid relinking later.
Search Console linking checklist
- Your Google login sees every property. DuckDive lists only what that login can open. In Google's permissions guide, only an owner can add you.
- Domain over prefix. Note which sites have a domain property and use it.
- One property per website. Decide now which property represents each site, so two DuckDive websites never fight over one.
- Admin rights in DuckDive. Only DuckDive admins can add sites or link sources; read-only members can ask questions but not configure anything.
- An AI app that accepts a remote address. Claude, Cursor and Claude Code need no extra check. If your SEO team works in ChatGPT, confirm the workspace tier on the ChatGPT page before anyone links a property.
-
Add DuckDive to your agent
From a terminal, the whole registration is one line:
claude mcp add --transport http duckdive https://api.duckdive.ai/mcp. The sign-in that follows, and the choice between a per-project and a per-user install, are on the Claude Code page. Chat apps take the addresshttps://api.duckdive.ai/mcpin their custom connector settings instead; those menus get renamed often, so search the app's settings for "Connectors", "Apps" or "MCP". -
Sign in with Google
A DuckDive page opens. Select Continue with Google. On the approval screen, leave the Also allow actions box checked so the agent can create sites and link sources, then select Allow. The sign-in page expires after 30 minutes, so finish it in one go.
-
Hand setup to the agent
Give it this one line. It reads your account state with
get_setup_statusand adds each site withadd_website, suggesting ones it finds in your Google account.
Paste into your agentSet up DuckDive for my website. Follow https://analytics.duckdive.ai/
agents.md -
Let it match each property
With
list_available_sourcesthe agent finds the Search Console site and GA4 property that clearly match each website, links them withconnect_website_sources, and asks you once about anything unclear. Linking starts the history import. -
Check the import, then ask
Most imports finish within minutes, longer for sites with deep history. Have the agent run
get_sync_status; when Search Console reads ready rather than importing or failing, ask your first question.
Ask your agentWhich of my DuckDive websites still have no Search Console site linked, and is the Search Console import finished on the ones that do?
Tools it calls list_websites get_sync_status
Search Console MCP questions, answered
Why don't the page rows add up to the site's total clicks?
Two reasons. The page list stops at 100 rows ranked by clicks, so the long tail sits outside it. And Google counts a site total by property but page rows by page; its Performance report help says chart totals "can sometimes differ" from table totals for that reason. Ask your agent to quote the site total from the summary rather than summing rows.
Do I need a Google Cloud project or a service account key?
No. You sign in with Google and your AI app connects to DuckDive over OAuth. There is no Cloud project to create, no JSON key to store, and no service account email to add to each Search Console property.
How far back can my agent see Search Console data?
The longest range the Search Console tools accept today is the last 12 months. Google itself keeps about 16 months of performance data, so nothing older than that can be imported by any tool. For a permanent archive, set up Google's bulk export to BigQuery.
Can I break queries down by page, country or device?
No. You get top queries and top pages, each up to 100 rows ranked by clicks, plus clicks, impressions, CTR and average position over time. For query-by-page, country or device splits, use Search Console or its API directly.
How many Search Console properties can I connect?
One site is free, with no credit card. Beyond that, pricing is per site. Each DuckDive website links one Search Console property, and your agent can ask about one site or loop across all of them in the same conversation.
Can my agent change anything in Search Console?
No. DuckDive has no tools for sitemaps, removals, URL inspection or settings, and the connector cannot delete data. Its only actions are setup inside DuckDive, such as adding a website or linking sources, and only for admins who allowed actions when connecting.
Does it work in ChatGPT, or only in Claude?
The three Search Console tools live on DuckDive's server, so a query report reads the same in Claude, Cursor, Claude Code or any other app that can add a remote MCP address with a Google sign-in. ChatGPT is the one to check before you start: DuckDive documents it only on team workspaces, and the ChatGPT setup page lists which tiers.
Sources
- Official MCP Servers for Marketing Teams, Fact-Checked, Digital Applied, checked November 16, 2026.
- ahonn/mcp-server-gsc, GitHub, checked November 16, 2026.
- AminForou/mcp-gsc, GitHub, checked November 16, 2026.
- Google Search Console Data Limitations, SEOTesting, checked November 16, 2026.
- Google Search Console Search Analytics API Can Now Access 16 Months of Data, Search Engine Journal (June 18, 2018), checked November 16, 2026.
- About bulk data export of Search Console data to BigQuery, Search Console Help, checked November 16, 2026.
- Best practices for managing service account keys, Google Cloud, checked November 16, 2026.
- Usage Limits, Search Console API, Google for Developers, checked November 16, 2026.
- Performance report (Search results): Overview and basic setup, Search Console Help, checked November 16, 2026.
- Search Analytics: query, Search Console API reference, Google for Developers, checked November 16, 2026.
- Add a website or platform property to Search Console, Search Console Help, checked November 16, 2026.
- Managing owners, users, and permissions, Search Console Help, checked November 16, 2026.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional.
