ga4-mcp-worker
Enables querying Google Analytics 4 data from Claude via a secure, read-only Cloudflare Workers MCP server with per-user OAuth sign-in, requiring no local installation.
README
ga4-mcp-worker
A read-only Google Analytics 4 MCP server that runs on Cloudflare Workers, so teammates can query GA4 from Claude with nothing installed locally — no Python, no gcloud, no ADC file, and as of this version, no shared password either. Adding it means clicking "Connect," signing in with your own Google account, and answering GA4 questions as yourself.
It mirrors the tool surface of the official
googleanalytics/google-analytics-mcp
stdio server, but calls the Google Analytics REST APIs directly, because Google's client libraries do
not run on the Workers runtime.
Read this first: docs/IMPLEMENTATION-NOTES.md § What has and has not been tested. Live end-to-end verification is in, but a few surfaces (the v1alpha endpoints, pagination) remain fixture-only.
Auth mode: per-user OAuth
Each teammate signs in with their own Google account. There is no shared secret and no single Google
credential that everyone relies on — see PRODUCTION.md § 1
and docs/PRODUCTION-PART2.md § 8 for
the full design and how it differs from the retired shared-secret mode. Sign-in is restricted to
@zuddl.com Google accounts (ALLOWED_EMAIL_DOMAIN in wrangler.toml).
GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET are still required — Google requires every app that
requests user data to be registered, and those two values are the registration, set once at deploy.
What is gone is the hand-minted refresh token and the password teammates used to type in.
Documentation
| File | What it is |
|---|---|
| SETUP-GUIDE.md (+ part 2) | First deploy, explained from scratch — assumes no prior OAuth knowledge. Creating the Google app, deploying, testing, connecting Claude. Start here. |
| PRODUCTION.md (+ part 2) | The identity model, monitoring, GA4 quotas, secret rotation, local dev, runbook, security checklist, OAuth internals. Read before a wide rollout. |
| docs/IMPLEMENTATION-NOTES.md | Gotchas this implementation gets right, what has and hasn't been tested, and running costs. |
src/index.ts |
The MCP server: tools, fetch helpers, normaliser, router. Heavily commented. |
src/google-oauth.ts |
The "sign in with Google" handler — /authorize and /callback. Heavily commented. |
wrangler.toml |
Worker config. Note nodejs_compat — see gotchas. |
.dev.vars.example |
Template for local dev. Copy to .dev.vars (gitignored). |
Our GA4 properties
| Property ID | Name | Measurement ID | Status |
|---|---|---|---|
314138239 |
New zuddl website GA4 property | G-JWBQ2Z84QF |
Canonical — use this one unless told otherwise. |
260479909 |
legacy | — | Legacy. Historical only; do not use for current reporting. |
328977581 |
legacy | — | Legacy. Historical only; do not use for current reporting. |
The server is property-agnostic: every tool takes property_id, and accepts either 314138239
or "properties/314138239". If omitted, it falls back to DEFAULT_PROPERTY_ID in wrangler.toml,
which is set to the canonical property so teammates do not have to memorise a number. Call
get_account_summaries to list everything the credentials can actually see.
Tools (9)
| Tool | Method + endpoint | What it is for |
|---|---|---|
run_report |
POST analyticsdata.googleapis.com/v1beta/properties/{id}:runReport |
The workhorse. Historical reporting. Full parameter set. |
run_realtime_report |
POST .../v1beta/properties/{id}:runRealtimeReport |
Last ~30 minutes. Separate, smaller schema. |
run_funnel_report |
POST .../v1alpha/properties/{id}:runFunnelReport |
Ordered step sequences and drop-off. |
run_conversions_report |
POST .../v1alpha/properties/{id}:runReport |
Conversions, ad spend, ROAS, attribution modelling. Restricted field list. |
get_custom_dimensions_and_metrics |
GET .../v1beta/properties/{id}/metadata |
Custom fields with the queryable apiName. Call before using custom fields. |
get_account_summaries |
GET analyticsadmin.googleapis.com/v1beta/accountSummaries |
Everything these credentials can read. No arguments. |
get_property_details |
GET .../v1beta/properties/{id} |
Timezone, currency, service level. Timezone explains date mismatches. |
list_property_annotations |
GET .../v1alpha/properties/{id}/reportingDataAnnotations |
Dated notes explaining spikes and dips. |
list_google_ads_links |
GET .../v1beta/properties/{id}/googleAdsLinks |
Linked Ads accounts. Confirms whether ad-cost metrics can have data. |
run_report accepts the full official parameter set: property_id, date_ranges (a list, so
period-over-period works in one request), dimensions, metrics, dimension_filter,
metric_filter, order_bys, limit, offset, currency_code, return_property_quota.
run_realtime_report takes the same shape minus date_ranges and currency_code.
Tool descriptions are long on purpose — they carry worked examples for every filter shape, and they are the only place the model learns the request format. Treat edits to them as behaviour changes.
Endpoints
| Endpoint | Auth | Purpose |
|---|---|---|
POST /mcp |
OAuth bearer token | The MCP connection (Streamable HTTP). There is no /sse. |
GET /authorize, POST /token, POST /register |
— | OAuth endpoints, implemented by @cloudflare/workers-oauth-provider. |
GET /callback |
— | Google redirects here after sign-in. Not for humans to open directly. |
GET /health |
none | Status check. 200 {"status":"ok","auth":"oauth", ...} or 503 + detail. See the note below — it means less here than it did in shared-secret mode. Leaks no data. Point an uptime monitor at it. |
GET / |
none | Plain-text status banner, names the current auth mode. |
Anything else returns 404.
/healthproves less in OAuth mode. There is no single shared credential to exercise, so a200only confirms the app registration is configured and the grant store (OAUTH_KV) is reachable — not that any particular person's sign-in is still valid. An individual user's grant can be expired or revoked while/healthis green. This is inherent to per-user auth, not a defect.
The Worker logs one JSON line per MCP request, per rejected auth attempt, per tool error, and per
health failure, plus OAuth-specific events (oauth_authorize_redirect, oauth_authorized,
oauth_domain_rejected, etc.). No token material is ever logged. All Google API calls send a custom
User-Agent (ga4-mcp-worker/1.0.0 (+cloudflare-workers)) so quota use is attributable.
Read-only enforcement
Three independent layers, all present:
- OAuth scope — every token, whether the old shared one or a teammate's own, is requested with
https://www.googleapis.com/auth/analytics.readonlyand nothing else (plusopenid/emailfor identity in OAuth mode — seesrc/google-oauth.ts, which grant no data access). Google rejects writes server-side. This is the real guarantee. SETUP-GUIDE.md walks through verifying it withtokeninfo. - Endpoint allowlist — anchored regexes checked inside the two fetch helpers before any network call. A future edit that introduces a write endpoint throws instead of dialling out.
- No mutating verbs — no
PATCH,PUTorDELETEanywhere in the codebase. There are exactly threefetchcall sites: the token mint (POST),gaGet(GET), andgaPost(POST).
:runReport, :runRealtimeReport and :runFunnelReport are HTTP POST but are queries — the body
carries the report definition because it is too large for a query string. Nothing in GA4 is created
or mutated.
Quick commands
npm install
npm run typecheck && npm run dry-run
npm run deploy
npm run tail
curl -s https://ga4-mcp-worker.YOUR-SUBDOMAIN.workers.dev/health
Set the two secrets both auth modes need (paste interactively — piping mangles values):
npx wrangler secret put GOOGLE_CLIENT_ID
npx wrangler secret put GOOGLE_CLIENT_SECRET
Connect from Claude Code — no header, since sign-in happens through the browser:
claude mcp add --transport http --scope user ga4 https://ga4-mcp-worker.YOUR-SUBDOMAIN.workers.dev/mcp
Then in an interactive claude session, run /mcp, pick ga4, and sign in with Google.
More
- Gotchas, test coverage, and running costs: docs/IMPLEMENTATION-NOTES.md
- First deploy walkthrough: SETUP-GUIDE.md → part 2
- Hardening, monitoring, and the OAuth runbook: PRODUCTION.md → part 2
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。