ad-preflight

A read-only Meta Ads, Google Ads and LinkedIn Ads MCP server and CLI that audits a campaign before you spend.

Can this campaign actually serve, and if it serves, can it convert?

ad-preflight reads a Meta, LinkedIn or Google Ads campaign from the platform's API and checks the fields that decide that. It also reconciles real spend across all three against what your records say. It's a CLI, a library and an MCP server, and it is read-only: it never activates, pauses or edits anything.

$ node examples/offline-demo.mjs        # a real preflight against a fake, broken campaign

LinkedIn campaign 123456789  "Q3 webinar - video"
status: ACTIVE

FAIL  Can win an auction: CPV / NONE / unitCost 0
      why: No optimization target and no manual bid is no bid at all. It shows ACTIVE and APPROVED and serves nothing.
      fix: Set an auto-bid target for the objective (e.g. MAX_LEAD, MAX_VIDEO_VIEW) or a manual unitCost above 0.
FAIL  Has somewhere to send people: webinar_15s: media post with no contentLandingPage and no contentCallToActionLabel
      why: Video and image posts without a landing page still serve and log clicks (profile, expand, play), but landing-page clicks stay at 0 forever. A placeholder form URN means no lead form.
      fix: Set contentLandingPage and contentCallToActionLabel on the post, or a real urn:li:adForm destination for lead gen.
PASS  Schedule lets it serve: ACTIVE, ends 2026-11-30
PASS  Creative is approved and serving: webinar_15s: ACTIVE serving
PASS  Geo resolves to the intended place: Columbus, Ohio, United States
PASS  Spend is bounded: total budget $1500.00, $50.00/day
SKIP  Landing page works: no click-out URLs (lead form or no destination)

DO NOT ACTIVATE: Can win an auction; Has somewhere to send people.
It is live now, spending up to $50.00/day while it can't do its job.

Try it with no credentials: node examples/offline-demo.mjs runs the real checks against mocked API responses for a campaign with two classic faults.

Why

A broken ad looks exactly like a working one in every dashboard summary: status ACTIVE, creative APPROVED, impressions rising. The difference is usually in one field that nobody checks. Each check here exists because that field was wrong on a real account:

• No bid. A LinkedIn campaign with optimizationTargetType: NONE and a $0 unit cost sat ACTIVE and APPROVED for three days and served one impression. The same config shipped again on the next campaign, because the "fix" was patched onto live campaigns and never reached the tool that created them.
• No destination. Every video ad an uploader built went out with no landing page. The ads served and logged clicks (profile views, video expands), so nothing looked wrong. About $2,000 bought around a dozen landing-page clicks and no leads, and for two months the copy got the blame.
• Wrong place. City targeting keys collide across states. A spec that says "Springfield" proves nothing about which Springfield the platform resolved.
• Wrong numbers. The record said a pilot was "paused, about $300 spent". The API said about $2,000. The gap stood for five weeks, and every decision made in that time used the wrong figure.

Install

npm install -g ad-preflight      # or use npx

Use a read-only token (ads_read, r_ads): the tool never writes, and a token that can't write can't be misused through it. Credentials come only from environment variables. Nothing is written to disk, and tokens travel in request headers, never URLs, so they stay out of platform error messages and logs. Errors are scrubbed of credential values before printing. Read-only scopes are enough. See .env.example.

Platform	Variables	Scope
Meta	META_ACCESS_TOKEN	ads_read
LinkedIn	LINKEDIN_ACCESS_TOKEN, optional LINKEDIN_AD_ACCOUNT_ID	r_ads, r_ads_reporting
Google Ads	GOOGLE_ADS_DEVELOPER_TOKEN, GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET, GOOGLE_ADS_REFRESH_TOKEN, optional GOOGLE_ADS_LOGIN_CUSTOMER_ID	adwords

API versions default to Meta v24.0, LinkedIn 202609 and Google Ads v25. LinkedIn retires monthly versions; if it answers "version is no longer active", set a newer one. Override them with META_API_VERSION, LINKEDIN_API_VERSION or GOOGLE_ADS_API_VERSION.

Preflight a campaign

ad-preflight meta <campaign_id> --region "Ohio"
ad-preflight linkedin <campaign_id> --account <ad_account_id> --country "United States"
ad-preflight google <customer_id> <campaign_id> --country US

Exit code 0 = SAFE TO ACTIVATE, 1 = DO NOT ACTIVATE, 2 = error. Add --json for machine output.

Check	Meta	LinkedIn	Google Ads
Schedule lets it serve	stop_time	status, runSchedule.end	serving status, end date
Bid can win an auction	cost/bid cap needs bid_amount	auto-bid target, or unitCost > 0	manual bids > 0, tCPA/tROAS targets set
Destination exists	link, CTA link, or an active lead form	contentLandingPage + CTA on media posts, article source, real urn:li:adForm	final URLs
Live	effective_status, review feedback	intendedStatus, review, isServing + hold reasons	approval and review status
Geo resolves where you meant	every city/region/zip's resolved region	every location URN resolved to its name	geo target canonical names
Spend is bounded	lifetime budget, end date, account spend_cap	totalBudget or end date	total budget or end date
Landing page works	fetched, status + redirects	fetched	fetched

Pass --region or --country whenever you know where the campaign should run. Without it, geo is listed for you to read but not verified.

Reconcile spend

ad-preflight spend --meta act_123 --linkedin 500000000 --google 123-456-7890 --days 30
ad-preflight spend --meta act_123 --ledger ledger.json          # exit 1 on any discrepancy

It pulls real spend per campaign from each platform's reporting endpoint, then reports:

• What can spend right now, as a headline number. On Google Ads, a paused campaign whose serving status is still SERVING is listed as armed: one switch from spending. A past end date (serving status ENDED) is the real brake.
• Discrepancies against your ledger: spend gaps over 2% (or $1), and campaigns your records call paused that can spend.
• Drift: live campaigns missing from the ledger, Meta accounts with no spend cap, and campaigns that spent money for zero clicks and zero leads (usually a structural break, so preflight them).

The ledger is what your records claim. See examples/ledger.example.json:

[{ "platform": "LinkedIn", "campaign": "123456789", "spend": 300, "status": "paused" }]

MCP server

claude mcp add ad-preflight -- npx -y ad-preflight mcp     # Claude Code

For Cursor, Claude Desktop and others:

{ "mcpServers": { "ad-preflight": { "command": "npx", "args": ["-y", "ad-preflight", "mcp"] } } }

Tools: preflight_meta_campaign, preflight_linkedin_campaign, preflight_google_ads_campaign, reconcile_ad_spend. All four are annotated read-only.

Claude Code plugin

/plugin marketplace add GroMarketing/ad-preflight
/plugin install ad-preflight@ad-preflight

It installs the MCP server and a skill that runs a preflight before any activation and reconciles spend before any CAC figure is quoted. Results are reported the same way every time: failures first, observed values instead of "looks fine", and an explicit verdict.

Library

import { preflightMeta, reconcile, formatReport } from 'ad-preflight';

const report = await preflightMeta('120000000000000000', { region: 'Ohio' });
console.log(formatReport(report));

What it doesn't do

• It doesn't judge creative, copy or targeting strategy. It answers whether the campaign is structurally able to serve and convert.
• It doesn't verify conversion tracking end to end. Whether your pixel or CAPI event fires on the landing page needs a browser test.
• It never writes. Fixes are printed as instructions; making them stays your call.

License

MIT