google-data-mcp

An MCP server for four Google data surfaces that answer to plain HTTP — no API key, no browser, no scraping service:

Read this before you install

Everything leaves from your machine's IP, and Google's per-IP budget accumulates over hours. This is not a caveat buried at the bottom; it is the main thing that decides whether this server is right for your job.

Measured: an unpaced burst was rate-limited at request 93, and after roughly 130 paced requests spread over hours the same address was refused on the first request of a fresh session. While writing this server, both the author's IPs — a home connection and a datacenter VPS — ended the day refused by Google Trends.

So this server caches every result on disk (~/.cache/google-data-mcp) and paces its requests. That is enough for interactive use: asking an agent a handful of questions, exploring a topic, checking a competitor. It is not enough for bulk work, and no amount of local code can make one IP behave like many.

If you need volume, the same clients run behind rotating proxies as Apify Actors: Google Trends · YouTube · Ads Transparency.

What is deliberately missing

YouTube transcripts. They look available — the watch page still lists caption tracks — but api/timedtext returns zero bytes and /youtubei/v1/get_transcript answers Precondition check failed even when sent the page's own INNERTUBE_CONTEXT and visitorData, and the ANDROID and IOS player clients are refused the same way. A transcript tool here would be a promise this server cannot keep, so there isn't one.

Install

pipx install google-data-mcp
# or: pip install google-data-mcp
# or, without installing anything: uvx google-data-mcp

Then register it with your MCP client. For Claude Code:

claude mcp add google-data -- google-data-mcp

Or by hand, in an MCP client config:

{
  "mcpServers": {
    "google-data": {
      "command": "google-data-mcp"
    }
  }
}

Notes on the data

• Trends values are relative within a single request. Two separate google_trends_interest calls are not comparable to each other; that is what google_trends_compare is for. Google caps a comparison at 5 terms and silently drops a 6th, so a 6th is refused rather than quietly ignored.
• Play reviews differ per locale. Each hl/gl pair returns a different set of reviews — six locales gave 120 unique reviews of the same app. Vary them to widen coverage rather than paging deeper in one language.
• Reviewer identity is omitted by default. The records carry an author name, avatar and Google account id; rating, text, date and version answer product questions without them. Set include_author if you genuinely need the name.
• Zero ads is a real answer. Google keeps advertisers whose ads have stopped running and still reports a count for them, so an empty ads list with a non-zero declared count is information, not a failure.
• searchVolume on trending terms is Google's own rounded bucket, not a precise count.

Legal

These are public pages, but none of them has an official public API and each platform's Terms of Service restrict automated access. You are responsible for how you use the output. No personal data is collected by default.

Development

python -m venv .venv && .venv/bin/pip install -e .
.venv/bin/python -m google_data_mcp        # speaks MCP over stdio

Each client is plain standard library and can be exercised on its own, which is the fastest way to check whether Google changed something:

.venv/bin/python -c "from google_data_mcp.play import PlayClient; print(len(PlayClient().reviews('com.spotify.music', limit=40)))"