Managing Google Ads campaigns in plain English 用一句自然语言,管理 Google Ads 广告
A Model Context Protocol (MCP) server that connects Claude Desktop directly to the Google Ads API: pause campaigns, move budgets, and adjust keyword bids by just asking. 一个把 Claude Desktop 直接接到 Google Ads API 的 MCP 服务器:暂停广告系列、调预算、改关键词出价,说一句话就行。
2026-07-07 · updated 2026-08-08 · first published on boheastill.com最初发表于 boheastill.com
The problem
Day-to-day Google Ads management is a loop of clicking through the same dashboard screens: check performance, pause what’s losing, raise the budget on what’s winning, adjust a bid. Each step is trivial, but the loop eats time, and the Ads UI buries simple actions under layers of navigation.
The constraint
Google’s Ads API is powerful but hostile to casual use: OAuth flows, developer tokens, customer IDs, YAML configs. The goal was to hide all of that behind natural language, without compromising on real API correctness, and to be demonstrable before any credentials existed.
The solution
An MCP server exposing seven tools to Claude Desktop:
- Campaigns. List with performance metrics (impressions, clicks, conversions, spend), toggle ENABLED/PAUSED, set daily budgets.
- Ad groups. List with status and parent campaign.
- Keywords and bidding. Pull keywords with metrics and current bids, add new keywords with match type and CPC limit, update max-CPC bids.
So “pause campaign 1002” or “raise the daily budget of campaign 1001 to $65” become things you say, not workflows you click through.
The design decision that sells it: mock mode
The server ships with a built-in simulation mode. With no credentials at all, it serves realistic campaign data, so the full experience is demonstrable in Claude Desktop in minutes, before touching a live ad account. Going live is then a config flip: an interactive OAuth helper script generates the token and writes the config for you.
This is a pattern I use deliberately: the client sees the working product first, and connects real credentials last. Trust flows from demonstration, not from promises.
Proof
Where else this applies
Any API your team touches daily through a clunky dashboard can become an MCP server: ad platforms, CRMs, e-commerce backends, internal admin panels. If your team already uses Claude, they can operate these systems by asking.
问题
日常管理 Google Ads,就是在同样几个后台页面里来回点:看效果、停掉亏钱的、给赚钱的加预算、微调出价。每一步都不难,但反复做很耗时间,而且 Ads 的界面把简单操作藏在好几层菜单里。
约束
Google Ads API 功能强大,但对偶尔用一下的人很不友好:OAuth 流程、开发者令牌、客户 ID、YAML 配置。目标是把这些都藏在自然语言后面,同时保证真实 API 调用的正确性,并且在还没有任何凭据时就能演示。
方案
向 Claude Desktop 暴露 7 个工具的 MCP 服务器:
- **广告系列。**列出效果指标(展示、点击、转化、花费),切换启用或暂停,设置日预算。
- **广告组。**列出状态和所属广告系列。
- **关键词与出价。**拉取关键词的指标和当前出价,按匹配类型和 CPC 上限新增关键词,更新最高 CPC。
于是“暂停 1002 号广告系列”“把 1001 的日预算提到 65 美元”,都变成说出来的事,而不是一层层点出来的流程。
真正打动客户的设计:Mock 模式
服务器内置模拟模式:完全没有凭据时,也能返回逼真的广告数据,几分钟就能在 Claude Desktop 里体验完整功能,不用碰任何真实广告账户。正式上线只需切换一个配置:交互式 OAuth 辅助脚本会生成令牌并写好配置。
这是我刻意坚持的模式:客户先看到能跑的产品,最后才接入真实凭据。信任来自演示,而不是承诺。
证明
还能用在哪
任何你的团队每天都要通过笨重后台去碰的 API,都可以做成 MCP 服务器:广告平台、CRM、电商后台、内部管理面板。如果团队已经在用 Claude,以后这些系统开口问一句就能操作。