Camberstack: a Google Ads MCP server

Connect Google Ads (and, optionally, Search Console) to Claude, ChatGPT, Cursor or any MCP client. Your AI reads the account with read-only tools and can propose changes, from negative keywords to whole Search, Performance Max, Display and YouTube campaigns. Nothing changes in the account until you approve a specific proposal, and every applied change can be undone in one call, on every plan. Risky changes (turning a campaign on, more than doubling a budget or bid, switching bid strategy) are flagged in the proposal before you approve. A demo account lets anyone try every tool first.

Hosted: https://camberstack.io/mcp (Free: every tool and unlimited changes on 1 Google Ads account at a time, with one switch a month; from November 16, Free covers one Google service at a time: Google Ads, Search Console or Google Analytics. Plus $7/mo: all three together on 1 account; Pro $15/mo: all three on up to 5 accounts; Agency $49/mo: up to 25).

Setup guides, with screenshots and troubleshooting: Connect Google Ads to Claude · Connect Google Ads to ChatGPT · every tool · Cursor, VS Code and others

Tools

tool	writes?	what it does
list_accounts	no	every account the login reaches, including client accounts under a manager (MCC)
account_overview	no	spend, conversions, cost per conversion per campaign
keyword_ideas	no	Keyword Planner ideas from seed keywords and/or a URL: monthly searches, competition, top-of-page bid range; marks ideas the account already targets or blocks
keyword_metrics	no	Keyword Planner numbers for an exact list, with the last 12 months of searches
run_gaql	no	any read-only GAQL SELECT
search_console_sites	no	the websites (Search Console properties) the login can read
search_console_summary	no	site-wide clicks, impressions, CTR and position for a window vs. the window before it
search_console_performance	no	organic clicks, impressions, CTR and position by query, page, country, device or date; text or RE2 regex filters
search_console_trend	no	a window vs. the one before it, per query and per page: rising, falling, new, lost, and rising searches losing position
search_console_opportunities	no	queries on positions 4-20 with projected click gain, top-5 queries with weak CTR, question-shaped queries
search_console_inspect_urls	no	URL Inspection for up to 20 pages: indexed or why not, last crawl, canonical chosen vs declared, structured-data errors
search_console_sitemaps	no	submitted sitemaps: last downloaded, URL count, errors and warnings
paid_organic_overlap	no	joins an Ads account's search terms with a site's organic queries: searches paid for that already rank near the top, and page-2+ searches with no ad
analytics_properties	no	the Google Analytics 4 properties the login can read; if Analytics isn't granted yet, a personal link that adds it
analytics_setup	no	a property's web streams (measurement IDs), key events and Google Ads links

The Search Console tools take either days or exact start_date/end_date, and fresh: true to include the last 2-3 days Google is still filling in. Requesting indexing and submitting sitemaps aren't available: Camberstack only holds the read-only scope.

The Google Analytics tools need analytics.readonly and analytics.edit (Google-verified sensitive scopes). Sign-in asks for them alongside Ads and Search Console; a connection made before they were added gets a personal /connect/analytics link on its first Analytics call, which adds both to the same Google account. The Cloud project needs the Google Analytics Data API and Admin API enabled.

What a proposal can do:

• Control: add negative keywords; pause or enable a campaign, ad group, keyword or ad; set a campaign's daily budget (never a shared budget) or its end date.

• Build, always created paused (building and spending are two separate approvals):

  • a Search campaign: ad groups, keywords, responsive search ads, targeting, negatives, sitelinks;
  • a Performance Max campaign: asset groups of text, images and YouTube videos, with the business name and logo on the campaign as Google's brand guidelines require (no shopping feed);
  • a Display campaign: responsive display ads;
  • a Demand Gen campaign with YouTube video ads, the API's only way onto YouTube (it refuses to create Video campaigns: MUTATE_NOT_ALLOWED, v23).

  Images are existing image assets or https URLs, which are downloaded SSRF-safe (src/fetch-safe.ts) and uploaded. A proposal can also add an ad group, keywords, an ad or sitelinks to an existing campaign.

• Tune: keyword and ad group bids, bid strategy, location and language targeting, presence-only location mode, final URL suffix, and a campaign's EU political advertising declaration (Google refuses every change to an account while any campaign lacks one).

• Measure: create a conversion action, set its counting, import a GA4 key event, scope a campaign to one conversion goal.

Risky changes carry a ⚠ line in the proposal summary: enabling a campaign, more than doubling a budget or a bid, switching bid strategy, removing the last location, a new primary conversion action. Undo removes only what Camberstack itself created; nothing the user built is ever deleted. (Undoing a GA4 import removes the action, since Google can't hide it again; importing again restores it.) Search Console is read-only.

How auth works

Camberstack is an OAuth 2.1 authorization server for MCP clients (dynamic client registration, PKCE). Its /authorize sends the user to Google for openid email plus adwords (Google Ads), webmasters.readonly (Search Console) and analytics.readonly + analytics.edit (Google Analytics). Any one is enough to connect ("à la carte"), Google Ads included: a tool that needs a missing one returns a personal /connect/<google-ads|search-console|analytics> link that adds just that service to the same Google account. The Google refresh token is stored AES-256-GCM encrypted and never leaves the server. The MCP client gets Camberstack's own tokens, which are stored only as SHA-256 hashes.

Two MCP SDKs, on purpose: /mcp is served by SDK v2 (@modelcontextprotocol/server + /node), which answers both 2026-07-28 clients (server/discover) and 2025-era ones (initialize, kept on a JSON-response transport). The authorization server (mcpAuthRouter, requireBearerAuth) stays on @modelcontextprotocol/sdk 1.x, because v2 ships only the resource-server half. Don't drop 1.x until v2 (or a companion package) has /authorize, /token and /register.

Run it

npm ci && npm test
ENCRYPTION_KEY=$(openssl rand -hex 32) \
GOOGLE_CLIENT_ID=… GOOGLE_CLIENT_SECRET=… GOOGLE_ADS_DEVELOPER_TOKEN=… \
BASE_URL=https://your.host npm run dev

env
ENCRYPTION_KEY	32 bytes, hex or base64. Losing it only forces users to reconnect; leaking it exposes stored Google tokens
GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET	OAuth client from a dedicated Google Cloud project with Google Ads API access. Redirect URI: <BASE_URL>/oauth/google/callback. Unset = pages only, sign-in reports "not open yet"
GOOGLE_ADS_DEVELOPER_TOKEN	sent when set
BASE_URL	public origin, the OAuth issuer (default https://camberstack.io)
DATA_DIR	SQLite location (default ./data)
FREE_ACCOUNTS / PRO_ACCOUNTS / PRO_EMAILS	Google Ads accounts each plan covers at a time, counted as distinct accounts used in the last 30 days plus one switch per 30 days, manager and demo accounts excluded (defaults 1 / 5) / emails treated as Pro
FREE_ONE_SERVICE	Free covers one Google service at a time (Ads, Search Console or GA4: the most recently used, over 30 days); paid plans cover all three. YYYY-MM-DD = announced until that day, 08:00 UTC (midnight Pacific): the site, billing, the account page and the welcome email describe the change, nothing is gated; from that day it's on, and only calls since then count, so every Free user picks a service. 1 = on now. Off unless set. The demo never counts
STRIPE_SECRET_KEY / STRIPE_PRO_PRICE_ID	Pro subscription via Stripe Checkout; unset = no upgrade links
STRIPE_AGENCY_PRICE_ID / AGENCY_ACCOUNTS	the Agency plan's price (its own product) / accounts it covers (default 25); unset price = no Agency offered. Which plan a user is on follows their subscription's price, so a Pro ↔ Agency switch in Stripe's billing portal (which must list both prices) moves it within the hour

The hosted HTTP server is npm run start:http (the Docker image runs node dist/server.js). npm start is a stdio mode that needs no env, database or Google login: it lists every tool for directory inspectors (Glama runs mcp-proxy -- pnpm run start), and any tool call answers with the hosted URL to connect to. pnpm-workspace.yaml turns off native builds for pnpm installs, so stdio mode installs anywhere; production installs with npm.

scripts/try-account.ts runs the read tools against a real account from an existing refresh token, skipping OAuth. Use it for dogfooding.

License

MIT