MCP
Give Claude live ads from any advertiser.
An ad library MCP for Claude is not a screenshot tool or a campaign manager. It is a way to load live advertising into your AI as structured, reasoning-ready context. Instead of asking Claude to guess what a market is running, or showing it one video at a time through vision, you connect Upspring's MCP once and Claude can query the live creative of any advertiser in a category.
This is loadable context, not files. The expensive processing work (video transcription, frame analysis, landing page extraction, advertiser enrichment) happens once when the ad is collected. What Claude receives is a compact, structured record: the hook, the claim, the offer, the format, and the destination, alongside independent scale grades for the advertiser. No screenshots, no vision calls, no guessing.
For the technical details on how raw ads become loadable context, see Rendering an ad into tokens. The short version: structured creative is typically around 10x fewer tokens than pushing raw video and pages through a model every time you ask a question.
Claude cannot watch 100 Meta videos.
Claude does not have native access to ad platforms. If you want it to analyze 100 video ads from a category, the default path is screenshotting each ad from the Meta Ad Library and uploading the images. That works for a portfolio review of three or four ads. It does not work at category scale.
Every screenshot requires vision model calls to parse what is in the image. That cost compounds fast: 100 ads is 100 vision calls, and the token count from raw image data blows the context window before you finish uploading. The model also cannot filter, compare, or retrieve by creative attributes (hook style, claim type, offer structure) it cannot see until it has already paid to process each image.
Video makes this worse. Claude cannot watch video. A video ad must be decoded, transcribed, have frames sampled and described, and on-screen text extracted before an AI can reason about it. Do that work every time someone asks a question about the category and you have built a system that cannot scale past a handful of advertisers.
The fix is to preprocess each video once at collection time (transcript, hook, claim, offer, on-screen text, format) into a compact, structured record, then retrieve that record many times. A processed ad is loadable context: Claude receives the hook, the claim, the offer, and the transcript as structured fields, not a video file it has to decode every query. Retrieval is cheap, comparison is fast, and you can ask questions across 100 ads, or 1,000, without re-running the expensive work.
Connect Upspring to Claude, ChatGPT, or Cursor.
Upspring is the official Claude Connector for advertising intelligence. Connecting it takes one setup step, then every question after that is natural language.
Claude Desktop or Claude Code: Open settings, go to MCP integrations, and add Upspring from the directory. If you are connecting a custom MCP client or configuring manually, the connection details are available in your Upspring account settings once you are logged in.
ChatGPT: In ChatGPT settings, add a custom action or GPT that connects to the Upspring API. Authentication is handled through an API key available in your Upspring account.
Cursor or other MCP clients: Add the Upspring MCP server through your client's integration settings. Configuration follows the standard MCP protocol; connection details are in your Upspring account.
Once connected, you do not interact with Upspring directly. You ask Claude or ChatGPT your question, and the AI decides when to query the MCP for the context it needs. The data comes back as structured JSON; the AI turns that into an answer.
For a step-by-step walkthrough of connecting the MCP and the kinds of questions it unlocks, see Bring the whole market into Claude, not just your own ad account. For details on what Upspring covers and how it differs from other data sources, see the MCP page.
What you can ask without uploading files.
Once connected, the questions get specific. Here are three examples of what you can ask Claude or ChatGPT through the Upspring MCP.
"Analyze 100 video ads in the project management category. What hooks repeat?"
Claude retrieves 100 video ads from that industry as structured records (each with hook, claim, offer, transcript, format already extracted), groups them by hook style, and returns the patterns that appear most frequently. This happens without uploading files, without screenshotting the Ad Library, and without re-transcribing video. The preprocessing was already done; Claude just retrieves and compares the records.
"Show me the longest-running video ads in this category and tell me what they have in common."
Claude retrieves ads marked as video, filters by observation span (how long each ad has been live), and analyzes the structured creative records for shared patterns: format, hook style, pacing, offers. Because each ad is already processed into reasoning-ready context, the comparison happens without vision calls.
"What hooks appear frequently in competitor ads but not in our own creative?"
Claude queries Upspring for live ads from advertisers in your space, retrieves the hooks, compares them against your own ad creative if you have connected your account, and returns patterns. This is competitor research, one of many use cases. Others include category research, brand tracking, and creative pattern analysis across any set of advertisers.
Your account vs the market.
There are three common categories of ads tooling for AI, and Upspring is not the first two.
Account MCPs (Meta's official ads MCP, Windsor, Porter) connect an AI to your Ads Manager so it can read your campaigns, manage budgets, and act on your behalf. They load your account data and performance, not the market.
Creative analytics platforms (Motion, Uplifted, GoMarble) analyze and tag your own video ads with hooks, scenes, and performance attribution. They process your creative, not what others are running.
Advertising intelligence MCPs (Upspring) load live ads from any advertiser as structured context. You query the market (a category, a competitor, a brand you are tracking), and the AI retrieves the ads, claims, offers, and landing pages of whoever you asked about. It does not connect to your Ads Manager. It does not tag your videos. It loads the live creative of the market as reasoning-ready context so your AI can answer questions about what others are running, which hooks dominate, and what patterns work.
Most teams use more than one. An account MCP manages execution. A creative analytics tool tags your library. An advertising intelligence MCP gives your AI the market context those decisions happen in. For more on the account vs market split, see Ads-account MCP vs Upspring.
Frequently asked questions
What is an ads MCP for Claude?
An ads MCP (Model Context Protocol integration) for Claude lets your AI load live advertising data as structured context without screenshots or manual exports. Upspring's MCP gives Claude access to ads from any advertiser, processed creative (hooks, claims, video transcripts), advertiser scale grades, and market trends across Meta, TikTok, YouTube, Google, LinkedIn, Reddit, and AppLovin.
Why not just screenshot ads from the Ad Library?
Claude cannot analyze 100 Meta videos from screenshots. The token count from raw images blows the context window, and every screenshot requires vision model calls to parse. An MCP preprocesses ads once at collection time (transcript, hook, claim, offer, format) into compact, structured records. Retrieval is cheap, the data is filterable by creative attributes, and you can ask questions across 100 ads, or an entire category, without uploading files.
Is Upspring a creative analytics tool like Motion or Uplifted?
No. Creative analytics platforms (Motion, Uplifted, GoMarble) analyze and tag your own videos. Upspring loads live ads from any advertiser across a market. It is not connected to your Ads Manager. It does not tag your library. It gives your AI the market (competitors, a category, brands you are tracking) as structured, reasoning-ready context.
What is the difference between an account MCP and an advertising intelligence MCP?
Account MCPs (Meta's official ads MCP, Windsor, Porter) connect your AI to your Ads Manager to read performance and manage campaigns. Advertising intelligence MCPs (Upspring) connect your AI to live ads from any advertiser, so it can see what others are running. Most teams use both: one for execution, one for market context.
Do I need to be technical to connect an MCP?
No. If you can use Claude or ChatGPT, you can connect an MCP. For Claude, add Upspring from the integrations directory. For ChatGPT, add it as a custom action. Once connected, you ask questions in natural language. The AI handles the queries.
What platforms does Upspring's MCP cover?
Upspring collects and processes advertising from Meta (Facebook, Instagram), TikTok, YouTube, Google (search, display, Shopping), LinkedIn, Reddit, and AppLovin. Coverage is live, meaning what Claude retrieves is what advertisers are running now, not a static snapshot.
Related reading: For the rationale behind treating ads as loadable context, see Ads as loadable context. For the technical preprocessing pipeline, see Rendering an ad into tokens. For Upspring's MCP in practice, see Bring the whole market into Claude, not just your own ad account.