To use Claude for media buying, connect your ad data to it through an MCP server. The platforms' own servers (Meta, Amazon Ads, Google Ads) each cover one platform and hand Claude raw rows; Peachblue's covers Meta, TikTok, Google Ads, and Amazon DSP in one connection, with every creative already grouped, scored, and tagged. Without a connection, Claude is a smart intern with amnesia. With one, it becomes something closer to an analyst who has read every row of your ad history. This guide covers both modes honestly, and how to set up the second one.
Disclosure: we make Peachblue, one of the tools below. The workflow advice stands regardless of whose tool you use.
What Claude does well with no setup at all
Plenty, actually. Paste in your creative copy and Claude critiques hooks and angles competently. Paste a screenshot of a results table and it reads the numbers. Describe your product and audience and it drafts briefs, generates angle lists, and rewrites headlines in volume. For creative ideation, it is already better than most brainstorm meetings.
The ceiling is just as real:
- It cannot see your accounts. Every number arrives by copy-paste, which means stale snapshots, tiny samples, and no follow-up questions the data could answer.
- Manual data entry invites errors. A pasted table loses the context around it: attribution window, date range, currency, which campaigns were excluded.
- It will reason confidently about numbers it half-remembers. Ask a follow-up two screens later and the model may be working from a garbled version of what you pasted.
Copy-paste Claude is a writing tool. The unlock is making it an analysis tool.
The unlock: MCP
The Model Context Protocol is an open standard that lets Claude call tools that fetch live data. Instead of you pasting numbers in, Claude queries the source and reasons over what comes back. And 2026 is the year the ad platforms embraced it: Meta, Amazon, and Google now run official MCP servers, and connector companies cover the rest. Which sounds like the problem is solved, until you actually try to run them together.
The DIY route: connecting each ad platform to Claude
Here is the current state of the official, platform-run MCP servers, and what each one gives Claude.
Meta ads MCP
Meta opened its official ads AI connectors in open beta on April 29, 2026. Connected, Claude can pull reporting for campaigns, ad sets, and ads, run Meta's own diagnostics (trends, anomalies, benchmarks), and create or edit campaigns, with new ads paused, for Meta only. What it does not do is judge the creative: no documented tool tags an ad by hook or format, or recognizes the same video uploaded twice. Setup, limits, and a full creative workflow are in Claude for Meta ads.
Google Ads MCP
Google open-sourced its official Google Ads MCP server in October 2025, giving LLMs direct access to the Google Ads API. You run it yourself, on your machine or your own cloud, with a Google Cloud project that has Google Ads API access and OAuth credentials (Google retired developer tokens in September 2026), and its tools only query: a search that runs Google Ads queries, resource metadata, and the list of accounts you can reach. Same shape: one platform, API-level rows, with Claude left to do the analysis in-chat.
Amazon Ads MCP
Amazon released its Ads MCP server in open beta in February 2026, a connection layer for AI agents into Amazon Ads, sponsored ads and Amazon DSP alike, and added a lighter three-tool version at unBoxed 2026. Connecting it takes an approved Amazon Ads API developer app, it can change campaigns as well as read them, and the data comes back as report rows: no flight pacing math, no creative-level analysis. Setup and a weekly DSP workflow are in Amazon Ads MCP in Claude.
TikTok ads
TikTok hosts its own as well. The TikTok for Business MCP Server needs no developer credential and comes in two versions, one that loads around 400 tools at once and one that starts with around 40 core tools. Its tools include creating campaigns, pausing ad groups and changing budgets, so it writes as well as reads; setup and limits are in TikTok Ads MCP in Claude. Connector services such as Pipeboard also run hosted TikTok servers. Either way, Claude gets the same kind of raw account data as from the platform servers above.
Or a multi-platform pipe
Connector services cut the number of connections. Pipeboard runs hosted servers for Meta, Google Ads, TikTok, Snap, and Reddit, and Windsor.ai connects 350+ sources, Amazon DSP among them. But they deliver raw rows too, so the stitching moves rather than disappears. All twelve options (official servers, pipes, and analysis layers) are compared in the best MCP servers for ad data.
What you end up with
Wire up everything above and this is the stack: three or four separate connections, each with its own auth flow, toolset, and vocabulary, each seeing only its own silo. Claude has to stitch the platforms together by hand in every conversation, reconcile attribution windows that do not match (Amazon DSP runs its own last-touch click and view model; social platforms differ), and reconstruct basic concepts like "this is the same creative running in nine ads" from scratch, every single time, because raw rows carry no creative identity, no scoring, and no tags. It works, in the way that four spreadsheets and a prayer work.
Or: one connection, with the analysis already done
Peachblue's MCP sits on top of synced, analyzed data across all four platforms instead, and the practical differences stack up fast:
| You ask Claude | With a raw platform MCP | With Peachblue's MCP |
|---|---|---|
| "How did last week go, and what drove the change?" | Metric deltas with no creative context; Claude guesses at causes | Performance summaries and trends tied to the specific creatives that drove them, so the "why" has a name and a thumbnail |
| "What is my best creative?" | Best ad per metric. The same creative split across 12 ads looks like 12 mediocre ads | Creatives grouped by fingerprint across every ad and platform, ranked by composite score |
| "Which hook styles are winning?" | No such concept exists in the data | Every creative is tagged across 31 dimensions; hook style is a filter, not a research project |
| "Compare Meta vs TikTok performance" | Each official connector sees one platform; a pipe returns both as unmatched rows | One connection covers Meta, TikTok, Google Ads, and Amazon DSP, with creatives matched across them |
| "How are my DSP flights pacing?" | Amazon connectors return raw report rows; no flight math | Budget, expected spend, pace percent, and CPM goals per order, computed the way an agency reports them |
| "Rank last month and tell me what to scale" | Claude invents a ranking method per conversation; answers drift between chats | Same composite score the app uses; Claude's answer matches your team's dashboard |
| "What do my winners have in common?" | Claude would need to inspect every creative itself, which does not scale in chat | Pattern detection and archetypes are precomputed; Claude retrieves and reasons |
| "What should we make next to keep scaling?" | No concept of a winner beyond a metric sort | Winning patterns across 31 dimensions feed the next brief, so scaling means repeating what your data proves works |
| "What is Reddit saying about my brand?" | Not ad-platform data at all | Brand Intel mentions, sentiment, and the editorial brief are tools Claude can call |
The thread running through every row: answers stay grounded in your performance data and tied to the creative level, which is where scaling decisions actually live. The loop this enables is the one that matters: find the winners, understand why they win, make more of what wins, and verify the next batch against the same score.
Two quieter advantages compound over time. History: Peachblue keeps what it syncs, so analysis is not capped by platform retention limits (Amazon's older DSP report API, for instance, keeps 95 days; your Peachblue history builds forward past it). Context efficiency: tools return ranked, aggregated payloads instead of raw row dumps, so Claude spends its context reasoning rather than arithmetic, which in practice means fewer wrong numbers.
To be fair to the raw option: platform-native MCPs are the right tool for platform-native lookups, and nothing stops you from connecting them alongside Peachblue. The distinction is what you want Claude to be. Raw rows make it a query runner. An intelligence layer makes it an analyst.
Peachblue's server exposes the same 23 tools that power our in-app agent, over your Meta, TikTok, Google Ads, and Amazon DSP data. Everything is read-only; nothing can touch your campaigns.
Setting it up
Prerequisites: a Peachblue account (Pro or higher) with at least one platform synced, and Claude Desktop or Claude Code.
- In Peachblue, open Settings and go to the MCP tab. Copy your connection URL.
- In Claude Desktop: Customize, then Connectors, then the plus button, then Add custom connector. Paste the URL.
- Approve the OAuth sign-in with your Peachblue account when prompted. No API keys.
Full setup details, including Claude Code, are in the MCP docs, and the complete tool list is in the tool reference.
What to actually ask
The queries that earn their keep daily:
- "How did last week compare to the week before, and what drove the change?"
- "Rank my top 10 creatives by composite score this month and tell me what they have in common."
- "Compare my two hero videos head to head. Which one deserves the next budget increase?"
- "Which hook styles are winning in my account right now?"
- "Which headlines are underperforming across my Google RSAs?"
- "What is Reddit saying about my brand this week, and is there an angle worth testing?"
Then the compound move: because Claude holds the whole conversation, you can chain analysis into action. Pull the patterns behind your winners, then in the same thread have Claude draft five new concepts that fit those patterns. Analysis to brief in one sitting, grounded in your own data at every step.
What this replaces, and what it does not
It replaces the export-paste-reformat loop, the "quick question" that used to mean opening four dashboard tabs, and a decent share of weekly report assembly. It does not replace your analytics app for visual browsing, it does not buy or edit media (by design), and it does not replace judgment: Claude with your data is a sharper analyst, not a decision-maker.
Honest limitations: you need a Claude plan that supports connectors, and MCP access starts on Peachblue's Pro tier (pricing). Agency workspaces connect the same way, with every call scoped per client. Fair-use rate limits apply but normal analytical use never meets them.
Where this is heading
Every serious ads tool will expose MCP within a year; the question will shift from "can my AI see my ad data" to "how good is what it sees." Grouped, scored, analyzed data beats raw rows for the same reason a good analyst beats a CSV export. If you want to try it on your own accounts, start a trial, sync a platform, and connect Claude in the time it takes to make coffee. And if you are choosing a layer in the first place, the best MCP servers for ad data and the 2026 creative analytics roundup are the honest places to start.