Google Play Console MCP

An MCP server for Google Play Console — 169 tools across publishing, Android vitals, revenue, ASO and store assets, with a confirmation gate on every destructive write.

Connect Claude, Cursor, VS Code, Codex, Gemini CLI, Windsurf, or any Model Context Protocol client to your Google Play developer account.

Ship a release. Ask whether it is crashing. Find out what it earned. See what users searched to find it. Reply to a review. All of it from the same session, and none of the destructive parts without you saying yes first.

"Is the 4.2.1 rollout healthy?"
  -> vitals freshness checked, crash rate by version code, worst issue's stack
     trace, current rollout fraction — and a recommendation

"How much did we make in July, and where did the installs come from?"
  -> earnings per currency, acquisitions by traffic source, top search terms

"Ship 4.2.2 to production at 10%."
  -> a preview you have to approve before anything moves

Why this one

There are a dozen Google Play MCP servers. This is the only one that covers the whole surface, and the only one with a real test suite.

Most servers wrap the Android Publisher API and stop. That API cannot tell you your crash rate, cannot tell you what you earned, and cannot tell you what people typed to find your app — those live in two other places entirely.

The five planes

Each needs different credentials, and the server tells you exactly which are missing and how to fix it.

Install

One line. It registers the server with the agents you pick:

uvx google-play-console-mcp install

You get a checklist of the agents found on your machine — Claude Code, Claude Desktop, Cursor, VS Code, Codex, Gemini CLI, Windsurf, Zed, Cline, Goose, OpenCode and 11 more — and it writes the right config file, in the right shape, for each one. No JSON to paste, no path to look up.

# pick agents up front and skip the prompts
uvx google-play-console-mcp install -a claude-code -a cursor -y

# read-only, with credentials wired in
uvx google-play-console-mcp install --read-only --credentials ~/sa-key.json

# see the exact command without running it
uvx google-play-console-mcp install --dry-run

install keeps the server itself out of your way: uvx fetches it on demand, so there is nothing to upgrade by hand. To keep a permanent copy instead:

uv tool install google-play-console-mcp   # then: google-play-console-mcp install
pipx install google-play-console-mcp
pip  install google-play-console-mcp

The wiring is done by add-mcp (Apache-2.0), which tracks each agent's config format so this server does not have to. It runs under npx; if you have no Node.js, install prints the exact config block to paste instead of failing.

{
  "mcpServers": {
    "google-play-console": {
      "command": "uvx",
      "args": ["google-play-console-mcp"],
      "env": {
        "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/service-account.json",
        "GOOGLE_PLAY_MCP_REPORTS_BUCKET": "pubsite_prod_rev_0123456789"
      }
    }
  }
}

Credentials

1. Service account — Google Cloud Console → enable the Google Play Android Developer API → create a service account → download the JSON key.
2. Invite it in Play Console → Users and permissions.
3. Grant these, which are separate and independently forgettable:

The bulk-reports permission is account-level. An app-level grant does not reach the storage bucket, and this is the most common reason those tools 403.

Find the bucket in Play Console → Download reports → Statistics → Copy Cloud Storage URI. There is no API that discovers it.

Then ask the agent to run check_reports_access, or read the gplay://config resource — both report what is configured and what each missing piece needs.

Safety

This server can ship to production, delete every localized listing, and refund real money. So:

18 destructive tools are behind a two-step confirmation gate. The first call returns a preview — what changes, what it affects, how to undo it — and does nothing. Only a second call carrying that preview's token executes.

The token is an HMAC over a canonical fingerprint of the operation, keyed with a per-install secret the model never sees. It catches an invented token, a token from a different operation, a parameter changed between preview and confirmation, a reused token, and an expired one.

The threat model is not a hostile user. It is a language model that hallucinates confirmation tokens, replays old ones, adjusts a parameter between the preview and the confirmation, and will cheerfully report that you approved something you never saw.

halt_release is deliberately not gated. Stopping a bad rollout is the one operation that reduces blast radius, and demanding a second round-trip during an incident is itself the failure mode.

Production cannot go to 100% in one step. The staged-rollout policy caps the first step at 20%; reaching everyone requires the separately named widening tool. The ceiling can be tightened but not disabled.

Everything is audited — previewed, confirmed, rejected, executed, failed — to ~/.google-play-mcp/audit.jsonl, with secrets scrubbed.

GOOGLE_PLAY_MCP_READ_ONLY=1 refuses every write outright, for exploratory sessions.

See Write safety.

What it knows that you would otherwise learn the hard way

• Screenshots must have no alpha channel. The icon must have one. A plain "PNG" export is RGBA — right for the icon, rejected for everything else. preflight_store_image catches it before a sixty-image bake fails at upload.
• Play requires at least two screenshots per type per locale. A set of one is valid file by file and still blocks the release.
• A subscription purchase left unacknowledged for three days is automatically refunded and revoked by Google. Silent revenue loss.
• The API accepts a 50-character title; Play policy caps it at 30. It uploads fine and is rejected at review.
• Vitals data lags by hours to days. A seven-day query that returns five days reads as an improvement rather than a gap, so freshness is a first-class tool.
• Earnings for a month are not published until around the 5th of the next. Reporting "$0" before that reads as a collapse rather than an absence.
• Earnings are never summed across currencies — 1000 JPY plus 10 EUR is a wrong answer that looks like a right one.
• com.acme.free must not collect com.acme.freemium's revenue. Attribution stops at a package-segment boundary.

Tools

169 tools. Full reference in the docs.

Plus 2 prompts (publish_screenshots, release_health_check) and 4 resources.

Tested

1,198 tests · 100% line and branch coverage · 15 modules
ruff clean · mypy clean · every release gated at --cov-fail-under=100

No live API call in the suite. Report parsing is tested against real UTF-16 bytes, image preflight against real PNG and JPEG headers.

Out of scope

Screenshot generation — use appscreens or any other tool, then preflight and upload here.

Licence

Apache-2.0.

_{Keywords: google play console mcp, google play mcp server, android publisher api mcp, play developer api, model context protocol google play, android app release automation, android vitals mcp, play console api claude, aso mcp, google play earnings api, play store screenshots api, android release agent}