Google Search Console MCP Server

An MCP server that gives Claude and other AI agents real control over Google Search Console — list your properties, check whether a URL is indexed, find the pages Google left out, run URL inspection in bulk, submit sitemaps, request indexing, and pull search analytics, all from a conversation.

Status: beta

Built for SEO practitioners tired of checking index status and submitting sitemaps by hand, one property at a time, and for the AI agents that can do it for them.

shahidali.co/tools/gsc-mcp is the project page: the same install lines, the tool list, and what the quota rules actually are, written for people who have not opened this repo.

Add it to your agent

Claude Code, in one command:

claude mcp add gsc -- uvx gsc-indexer-mcp

Anything else that speaks MCP, by hand — this is the whole config:

{
  "mcpServers": {
    "gsc": { "command": "uvx", "args": ["gsc-indexer-mcp"] }
  }
}

uvx fetches and runs the published package with nothing installed first, which is why every button above is a one-liner. A .mcp.json with exactly that content sits in this repo, so cloning it is also enough.

No API key goes in any of these. Sign-in happens once, in your browser, when you first call gsc_setup() — see The OAuth client is handled for you.

Project status

Beta. The whole surface is wired up, and it has now met a real Google account — on one operating system, driven by one person.

0.1.0 drops the pre-release suffix so that pip install finds it without --pre. That is a decision about reach, not a claim of maturity: this project needs bug reports from people who are not its author, and a package pip refuses to install by default does not get them. Read the Known issues in CHANGELOG.md before pointing it at a property you care about — in particular, a transient rate-limit from Google currently ends the whole batch. Bug reports are the point; please file them.

Fifteen tools are registered on the server and covered by a wire-level smoke test that connects a real MCP client session and confirms every tool answers with a description. Storage, quota accounting, OAuth, and config are the foundation underneath them. Sign-in has been walked for real, on Windows against a live Google account — see docs/manual-smoke.md, which is also where the defects that run found are recorded. Submission has since driven a real browser and had Google confirm real Request Indexing calls, in Chrome and in Brave, against live properties. That is one platform and one operator, and every automated assertion about it is still made against a fake; see Known gaps for what that does and does not buy you.

Released versions and what changed in them are in CHANGELOG.md. Watch or star the repo if you want to know when the milestones above are verified on a platform other than Windows.

Tools

Shipped and registered on the MCP server today:

Why quota accounting is the hard part

Most tools in this space get Google's limits wrong, then get throttled and blame detection. Both limits that matter are per property, not per account:

Properties are independent, so eight properties means eight independent budgets. This server tracks slots individually rather than counting a daily total, so it knows the exact minute the next slot opens — and it deliberately over-counts rather than under-counts when a race is possible, because a short wait is cheaper than a hard Quota Exceeded.

Finding what is not indexed

gsc_find_unindexed collects candidate URLs — from the property's registered sitemaps, from URLs already in the local store, or both — inspects the ones whose last inspection has gone stale, and reports each unindexed URL with a reason. limit caps how many URLs are inspected, not how many come back: an inspection spends budget, and a cap that only trimmed the output would pay full price for an answer it discarded. Which URLs a capped run reaches follows the store's own URL ordering (alphabetical), not staleness, so a capped run is a sample rather than a worst-first sweep.

gsc_audit answers the same question from the store alone — no HTTP inspection, no budget spent. It is point-in-time: it reports what the last inspection found, and carries as_at and a stale count so you can tell how old that picture is. There are deliberately no movement numbers.

Reason codes

There are ten of them. submitting_helps on each row is the one to act on before calling gsc_request_indexing:

If you have read the eight-code list in the design notes and counted ten here, the extra two are discovered-not-indexed and unknown-to-google, kept separate from crawled-not-indexed on purpose. All three answer yes to "can a quota slot move this", but not to "what do I do first", and that is the part you act on. Discovered and unknown are pages Google has not judged — it has not fetched them yet, so submitting is the whole remedy. Crawled-not-indexed is a page it fetched and passed on; a fresh crawl can reverse that verdict, but resubmitting byte-identical content re-crawls to the same one. Improve the page, then submit it.

Slots are roughly eleven per property per rolling day and unrecoverable, and crawled-not-indexed is usually the biggest bucket on a real site. submitting_helps: true means the slot can work, not that today is the day to spend it.

A URL whose state could not be established — a failed inspection, or a result a re-check could not confirm — is reported as undetermined, never as unindexed. Absence of evidence is not a finding.

Submitting URLs

Google has no API for Request Indexing that a tool like this can use, so submission goes through your own browser: the bridge extension, loaded into the profile you paired during gsc_setup, clicks Request Indexing in a real Search Console session. That has three consequences worth knowing before you spend anything.

It is slow, and the slowness is the feature. Submissions are paced 130–180 seconds apart. That gap is not a placeholder and not a bug report — it is the interval proven not to draw a throttle over long runs. Five URLs is therefore up to fifteen minutes of wall clock, and the tool that does it blocks for all of it.

Quota is per property and small. Roughly eleven slots per property, on a rolling 24-hour window — each slot frees 24 hours and a minute after its own use, not at midnight. Properties are independent budgets. Call gsc_quota first and act on spendable_free, not free: spendable_free subtracts the daily_reserve you set aside in config. A spent slot is unrecoverable; there is no undo.

One run at a time. The bridge drives a single browser tab in your real profile and listens on one fixed local port, so a second run while one is going is refused outright rather than queued. That covers both directions: gsc_start_indexing_job while a job or a gsc_request_indexing call is in flight, and gsc_request_indexing while a job is running.

Which tool

What a run does when things go wrong

A run stops early rather than burning the rest of the batch against a server that is already refusing: a quota_exceeded, a captcha, a rate limit, or a signed-out session ends it. stopped_early and stop_reason in the result say so, and the URLs never attempted keep their slots. A job that ended this way lands in state stopped_throttled; one you stopped by hand lands in stopped_user.

URLs that could not be routed to any known property come back as no_property, and ones that found no spendable slot as no_quota. Neither reached the browser and neither cost anything — they are reported apart from failures on purpose, because the fixes are different: run gsc_list_sites for the first, wait for the second.

If the server is restarted while a job is running, that job's row is closed out as failed at the next startup, and any submission row left open is settled against the property's ledger. Nothing is silently resumed — a background worker does not survive the process that owns it.

Before your first submission

The extension must be loaded in the browser profile you paired, and the browser must be one this server can drive. gsc_setup walks both; gsc_doctor re-checks them. If the browser is closed, auto_launch_browser (on by default) opens it. The first run of all pairs the extension to the bridge, which needs the browser window in front of you.

Requirements

• Python 3.11 or newer
• A Google account with Search Console properties
• A Chromium browser — Chrome, Brave, Edge, Vivaldi, Opera or Chromium. gsc_setup loads a browser extension into one of your existing profiles, and there is no way to complete setup without one. Firefox and Safari are not Chromium and will not work.
• No Google Cloud setup. gsc_setup fetches the OAuth client this project ships. You can still supply your own — see Install below.

Install

The distribution is gsc-indexer-mcp. Note the indexer: plain gsc-mcp on PyPI is an unrelated project by another author, so installing that name gets you someone else's server. Everything inside this one is still gsc_mcp — the import package, the console script, the config directory.

pip install gsc-indexer-mcp

Versions up to and including 0.1.0a6 were pre-releases and needed --pre. From 0.1.0 they do not.

Or skip installing entirely and let uvx gsc-indexer-mcp fetch it on demand, which is what the buttons at the top of this page do.

You can also take the wheel from the latest release — that one differs in one way, described below: it carries the OAuth client, where the PyPI build downloads it at first setup.

Or from a checkout, which is what you want if you intend to change anything:

git clone https://github.com/Mrshahidali420/google-search-console-mcp.git
cd google-search-console-mcp
python -m venv .venv
.venv/Scripts/python -m pip install -e .   # POSIX: .venv/bin/python

The OAuth client is handled for you

You do not need a Google Cloud project. gsc_setup() finds a client in this order — environment variables, the client baked into a release wheel, then a one-time download of the client this project ships, cached in your config directory. A source checkout takes the third path; everything after the first run is offline.

The tracked source deliberately contains no client. A secret committed to a public repository is permanent in git history, and GitHub reports Google client secrets to Google, which can revoke them — breaking every user at once. So it is delivered as a release asset and a build-time injection instead; docs/google-cloud-oauth.md has the mechanism. To be plain about it: that asset is public and anyone can read it, exactly like the wheel. What it buys is rotation, not secrecy — an installed app's security comes from PKCE, which this uses.

Using your own OAuth client instead

Optional. Worth doing if you want the consent screen to carry your own app name, or your organisation requires its own client. Set GSC_MCP_CLIENT_ID and GSC_MCP_CLIENT_SECRET and they override everything above, in a release build or a source checkout alike.

All of this happens in Google Cloud Console, free, and takes about five minutes. Do the steps in order — step 4 is the one people skip, and skipping it fails at the very end.

1. Create a project. Use the project dropdown in the top bar → New project. Any name. Wait for it to be created, then make sure it is the project selected in that dropdown; everything below applies to the selected project only.

2. Enable the Search Console API. Search "Search Console API" in the console's search bar, open it, and click Enable. Without this, every call returns a 403 that mentions the API being disabled.

3. Configure the OAuth consent screen. In the left menu: APIs & Services → OAuth consent screen.

   • User type: External. ("Internal" only exists on Workspace accounts and restricts the app to your own organisation.)
   • App name: anything — you are the only user who will see it.
   • User support email and developer contact email: your own address.
   • You do not need to add scopes here. This server requests its scope at sign-in time, and adding it on this screen does not change what you are granted.
   • Save through to the end.

4. Publish the app — or add yourself as a Test user. Either works; publishing is better, and this is the step people skip.

   Every new app starts in Testing status, where Google lets only listed test users sign in. Everyone else gets 403: access_denied on the consent screen — after creating the project, enabling the API, creating the client, setting the environment variables and running gsc_setup. Nothing earlier warns you. A Testing app also expires its refresh token every 7 days, so sign-in breaks a week later in a way that looks like a bug here and is not.

   Both Search Console scopes are classed non-sensitive by Google, so publishing costs nothing: no verification, no security assessment, no user cap. Audience → Publishing status → Publish app, and both problems above disappear.

   If you would rather stay in Testing, find Test users (on newer consoles: Audience → Test users) and add the Google address that owns your Search Console properties. That works indefinitely for your own account, at the cost of signing in again every 7 days.

5. Create the OAuth client. APIs & Services → Credentials → Create credentials → OAuth client ID, application type Desktop app. Copy the client ID and client secret it shows you — the secret is shown once, though you can always create another client.

   Desktop app is the right type: this server signs you in over a loopback redirect on 127.0.0.1, which is what that type allows. Do not pick "Web application" and do not add a redirect URI by hand.

6. Set the two environment variables below before starting the server. Without them, any tool that needs to talk to Google returns {"ok": false, "error": "not_configured", ...} rather than doing anything — gsc_list_sites, gsc_check_status, gsc_performance, and gsc_submit_sitemaps all behave this way. gsc_doctor and gsc_quota are the two exceptions: gsc_doctor still runs and reports oauth_client: not ok as one line in its checks list rather than failing outright, which makes it the right first tool to run when something is stuck; gsc_quota is local-only and returns [] on an empty store regardless of OAuth configuration.

export GSC_MCP_CLIENT_ID="your-client-id.apps.googleusercontent.com"
export GSC_MCP_CLIENT_SECRET="your-client-secret"
# Windows PowerShell:
#   $env:GSC_MCP_CLIENT_ID = "your-client-id.apps.googleusercontent.com"
#   $env:GSC_MCP_CLIENT_SECRET = "your-client-secret"

Connect it to an MCP client

Point your MCP client at the installed console script. For Claude Desktop, add to its claude_desktop_config.json:

{
  "mcpServers": {
    "gsc-mcp": {
      "command": "C:\\path\\to\\google-search-console-mcp\\.venv\\Scripts\\gsc-mcp.exe",
      "env": {
        "GSC_MCP_CLIENT_ID": "your-client-id.apps.googleusercontent.com",
        "GSC_MCP_CLIENT_SECRET": "your-client-secret"
      }
    }
  }
}

On POSIX, use .venv/bin/gsc-mcp as the command instead. Any other MCP-speaking client that can launch a stdio server works the same way — the entry point is the gsc-mcp console script installed above, and the server talks the standard MCP stdio transport (gsc_mcp.server:main).

This has been exercised against a real Google account through a real MCP client session on Windows; it has not been run against Claude Desktop specifically, nor on macOS or Linux. See Known gaps.

Getting started

Once the server is installed and connected, the whole of setup is one tool called repeatedly. gsc_setup() is idempotent: it never resumes a session, it re-reads the whole state on every call, and it hands back the single next thing to do. Call it in a loop until it returns ok: true.

1. Install and connect — see Install above.

2. Nothing. The OAuth client is handled for you — gsc_setup() downloads it on the first run and caches it. Set GSC_MCP_CLIENT_ID and GSC_MCP_CLIENT_SECRET only if you want to use your own.

3. Run gsc_setup(). It opens a Google consent screen in your browser and returns the URL as well, so a headless machine can still complete it by hand. Approve it, then call gsc_setup() again — the second call collects the redirect, stores the token, and moves on. Nothing about your sign-in is returned to the caller: not the token, not the authorization code, not the PKCE verifier.

4. Load the bridge extension. gsc_setup() will tell you where it extracted the extension to and which browser profile to load it into. In that browser: open its extensions page (chrome://extensions, brave://extensions, edge://extensions, and so on — gsc_setup() gives you the exact URL for your browser), turn on Developer mode, choose Load unpacked, and select the folder it named.

You will see a warning banner, and it is expected. The extension asks for the debugger permission, so Chromium shows a prominent bar saying an extension is debugging your browser, and may warn you when you enable Developer mode. That permission is not incidental. Search Console applies a soft throttle to Request Indexing clicks that did not come from a real pointer, and synthetic DOM clicks trip it. The extension therefore issues trusted input events through the Chrome DevTools Protocol instead, which is what debugger grants and the only way to grant it. The banner is Chromium correctly reporting a real capability — read it as "yes, this is the extension you just installed", not as malware. It only ever attaches to search.google.com, the one host in its host_permissions.

5. Run gsc_doctor(). Seven checks, in order: oauth_client, token, config, store, properties, browser, extension. Every failing one carries a concrete fix. Sample output, on a machine where everything is working:

{
  "ok": true,
  "checks": [
    {"name": "oauth_client", "ok": true, "detail": "configured", "fix": ""},
    {"name": "token", "ok": true, "detail": "token file present", "fix": ""},
    {"name": "config", "ok": true, "detail": "config valid", "fix": ""},
    {"name": "store", "ok": true, "detail": "schema version 2", "fix": ""},
    {"name": "properties", "ok": true, "detail": "2 properties", "fix": ""},
    {"name": "browser", "ok": true,
     "detail": "Google Chrome / Default is the profile to use", "fix": ""},
    {"name": "extension", "ok": true,
     "detail": "the gsc-mcp bridge extension is installed in Google Chrome / Default at version 1.10.0; whether its background service worker is running is not checked in this milestone",
     "fix": ""}
  ]
}

Reading the browser and extension checks

These two are about your local machine rather than your Google account, and they are worded carefully because the failure modes are easy to misread.

"Could not be checked" never means "not installed." The extension check reads your browser's own preferences files. Those files are frequently locked, mid-write, cloud-synced, or held open by antivirus. When a read does not happen, the check says the question could not be checked and that the extension may already be there. That is not a polite way of saying it is missing, and the fix is to run gsc_doctor() again (closing the browser first if it is running) — not to reinstall an extension that is sitting right where you put it. The same distinction runs through gsc_detect_browsers, whose has_extension field is three-valued: true present, false every preferences file was read and it was not among them, null the check could not be performed.

Microsoft Edge can report a Microsoft account where a Google one is expected. Edge stores signed-in account addresses in the same file and the same key Chrome uses for Google accounts, but by default it signs profiles in to Microsoft identities. Nothing on disk tells the two apart. So for an Edge profile, an address found is not evidence of a Google sign-in — and if your Microsoft address happens to be the same as your Google one, what looks like a confirmed match is not confirmed at all. The tools hedge this rather than assert it: an Edge profile is reported with "this profile's Google sign-in could not be confirmed". Check it yourself before relying on it. Brave, Vivaldi, Opera and plain Chromium record no Google account at all and are hedged for the different reason that there is nothing to read.

A changed extension ID means re-pairing, not breakage. The extension ships with no manifest key, so Chromium derives its ID by hashing the absolute path it was loaded from. That ID is stable for as long as the extraction directory is stable — and it changes if that directory moves: an upgrade that relocates the config directory, a different GSC_MCP_HOME, a migration to a new machine. When it changes, the extension check stops recognising the loaded copy and reports it as not installed. Nothing is broken and nothing is corrupted; the fix is to load the unpacked extension again from the new folder, exactly as in step 4.

A green extension check means registered, not running. It says the extension is loaded into that profile at that version. Whether its MV3 background service worker is alive is a separate question, and only the bridge can answer it: at submission time it waits for a connection, wakes an evicted worker if none arrives, and fails with extension_not_connected if that does not work either. gsc_doctor still reports registration only.

Development

.venv/Scripts/python -m pip install -e ".[dev]"   # POSIX: .venv/bin/python
.venv/Scripts/python -m pytest -v

Architecture

gsc_core is a standalone engine with no MCP dependency, so it can be driven by the MCP server, a CLI, or a future desktop app without change.

Privacy

To tell you which browser profile to use, this tool reads the profile list and the signed-in account address out of the browser's own files on your machine — Local State and each profile's Preferences / Secure Preferences. The same files are read a second time, for a different reason, to find out whether the bridge extension is loaded in that profile and at what version.

That read is entirely local. The address is used in memory to show you which profile is signed into which account, and:

• it is never transmitted anywhere,
• it is never written to disk by this tool,
• it is never written to the log, at any level — failures reading these files are logged by exception type name only, precisely so that neither an address nor a path containing your Windows username can end up in a log file you might attach to a bug report.

Nothing in the tool opens these files for writing. No address is ever returned by a tool either, in either direction — not one found in a profile, and not your own authorised address. A tool result is rendered into a transcript and retained by whatever MCP client is driving the server, none of which this project controls, so a profile is identified by browser and profile directory and nothing else.

One path is the exception, and it is deliberate. When gsc_setup finds the bridge extension is not loaded yet, its result carries the directory the extension was unpacked to — on Windows, C:\Users\<your username>\AppData\Roaming\gsc-mcp\extension. On most machines that path contains your operating-system account name. It is returned because "Load unpacked" in Chrome asks you to pick that exact folder, and a set-up instruction you cannot follow is not a privacy win. It is the only path any tool returns, it appears only in the one step that needs it, and it is never written to the log. If your account name is something you would rather not have in a transcript, run that step, then clear the transcript.

Known gaps

Stated plainly, because they are the things a reviewer should look at first:

• No test proves icacls actually applied an ACL on Windows — the Windows test only observes that the call was made, so _harden could no-op there and the suite would stay green. The POSIX equivalents now execute on Linux and macOS on every push, so this gap is Windows-only.
• The submission path has now submitted real URLs, but the test suite still cannot prove it. As of 2026-08-05 it has driven a real browser, clicked Request Indexing, and had Google confirm the request — on Windows 11, in both Chrome and Brave, across several days and several properties, including running the quota ledger up against a real refusal from Google. What the automated suite exercises is still fakes only: every assertion about the extension bridge, the pacing, the quota reservation, the run loop and the job worker is made against a stand-in. So "it submits URLs" is now supported by a manual log rather than by CI, and the states a fake cannot reach remain the submission pass in docs/manual-smoke.md. Live running has found defects the suite missed at a steady rate; assume it has not run out of them.
• The sign-in path has now been walked for real — once, on one machine (Windows 11, Chrome, Python 3.13, 2026-08-04), against an account with nine properties. Steps 1-8 of docs/manual-smoke.md pass. One run on one platform is not coverage: it found two defects that the whole automated suite had missed, which is the argument for running it again on macOS and Linux rather than for trusting it.
• The doctor cannot detect a stale extension build, and no longer pretends to. Chromium records no manifest snapshot for an unpacked extension — and unpacked is the only way this bridge is installed — so the version the browser is actually running cannot be read from disk at all. The check now reports the version on disk and says the loaded one is unreadable, rather than passing the first off as the second. Closing this needs the extension to report chrome.runtime.getManifest().version over the bridge, which is a protocol change and is deferred until the bridge has been exercised against a real submission.
• The bridge port defaults to 8765, and two runs cannot share it. A second gsc-mcp process, or anything else already on that port, fails to bind — the run now fails within seconds naming the port and the remedy, rather than blaming the extension after the full connect timeout. Moving the port (bridge_port in config) requires setting the same port in the extension's options page; the two are paired, not discovered.
• macOS and Linux browser detection has only ever run against fixtures, never on real hardware, here or in CI. The first person to run the smoke checklist on a Mac or a Linux box is performing that test.
• gsc_detect_browsers reports matches_authorised_account as null on every real machine today. The flag reads an account_email key from the stored token, and nothing writes it: the current scope set returns no identity claim and the consent step does not record the authorising account. The plumbing is correct and inert. Treat the field as "unknown", not as "does not match".
• The extension check reports whether the extension is registered, not whether it is working. Only the bridge learns whether the MV3 worker is alive, and only at submission time.
• The published package is on PyPI as gsc-indexer-mcp, not gsc-mcp — that name belongs to an unrelated project by another author, and installing it gets you their server, not this one. The two distribution channels also differ in one substantive way: the GitHub release wheel embeds this project's OAuth client, while the PyPI wheel deliberately does not (a secret on a public index is a revoked secret), so a PyPI install downloads the client from the client release asset at first gsc_setup(). Both paths have been exercised, but they are not the same path, and a bug in one would not show up in the other.
• mcp is pinned >=2.0,<3. The server is written against mcp 2.0's MCPServer API — 1.x spelled it FastMCP and the two shapes are mutually exclusive, so this package now hard-conflicts with anything still requiring mcp<2 (the mirror image of the conflict the old <2.0 cap created).

Contributing

Contributions are genuinely welcome — issues, pull requests, bug reports, docs fixes, all of it. Start with CONTRIBUTING.md. Good first issues are the CI legs listed under Known gaps.

The single most useful thing you can contribute right now is a bug report from a machine that is not Windows. Everything here has been exercised by hand on one operating system, by one person; see Known gaps.

• Report a bug · Request a feature
• Security policy — please report anything involving tokens, OAuth or the local bridge privately
• Code of conduct

License

Functional Source License 1.1 with an Apache 2.0 future grant (FSL-1.1-ALv2).

In plain terms: use it, modify it, contribute to it, run it on client work — just don't sell a competing product built from it. Every release converts to Apache 2.0 two years after publication, so nothing here is locked away permanently.