Claude MCP Server for Ad Testing

Three steps

Key. Connect. Ask.

No hosting, no glue code. The server runs on your machine and talks to the AdTest.AI API with your key.

Get an API key

Sign in at app.adtest.ai, open Developer, and generate a key. It is tied to your wallet, so credits are billed per analysis.

Add the server

One command in Claude Code, or a six-line entry in Claude Desktop's config. Restart Claude and the analyze_advert tool appears.

Ask Claude

"Analyze this ad" with a URL, a file, or the copy itself. Claude runs the analysis and reads you the report, weakest dimension first.

Install in one line

Two minutes from key to first score.

The server runs locally through npx, so there is nothing to host. Your API key stays on your machine and is sent only to the AdTest.AI API.

Claude Code

$ claude mcp add adtest \
    -e ADTEST_API_KEY=adk_your_key \
    -- npx -y adtest-mcp

Claude Desktop, or any MCP client: add to claude_desktop_config.json

{
  "mcpServers": {
    "adtest": {
      "command": "npx",
      "args": ["-y", "adtest-mcp"],
      "env": { "ADTEST_API_KEY": "adk_your_key" }
    }
  }
}

Requires Node 18 or newer and an AdTest.AI account with credits. Get your key at app.adtest.ai under Developer.

claude · adtest mcp

# You
› Analyze this ad: https://yourbrand.com/spring-launch.mp4
  Brand: yourbrand.com

# Claude → analyze_advert(url, brand_url)
  ✓ job submitted · video · polling
  ✓ report ready in 2m 40s

# Claude
  Core message 4.6 /5 · Emotional resonance 4.4 /5
  Weakest: Brand distinctiveness 3.4 /5
  Recommendation: bring the logo in before the end card…

1 tool · analyze_advert image · video · copy AI-only scoring

What Claude gets back

The full report, as text Claude can reason over.

One tool, analyze_advert. Give it exactly one creative and it returns the same AI-only scoring report the AdTest.AI app produces, so Claude can compare variants, explain a weak score, or draft the fix.

url — a public image or video URL. AdTest downloads it.
file_path — a local image or video file. The server uploads it.
text — ad copy: a headline, a script, a caption.
brand_url (optional) — the advertiser site, for sharper brand and logo detection.
Scores across all 13 dimensions, what is working, and a recommendation per dimension.
Video is supported. Long files run as a background job and Claude waits for the result.
Charged once, on success: 1 credit per completed analysis. Failed jobs are never billed.

Works with other connectors

Close the loop in one conversation.

Your ad platform's connector can tell Claude which ad underperformed. AdTest tells Claude why. Ask for the lowest-CTR ad from last month, then say "run it through AdTest" and Claude returns the weakest dimensions and the edit to make, without leaving the chat.

Open source, MIT licence: github.com/menaker/adtest-mcp
Package: adtest-mcp on npm
Listed in the official MCP Registry as ai.adtest/adtest-mcp and on Glama
Audience validation with real people stays in the web app. The MCP is AI-only by design.

MCP FAQ

Quick answers on the MCP server.

Does the MCP server include audience validation with real people?

No. The MCP server runs AdTest.AI's AI-only scoring. Validating a winner with a real first-party audience is done in the web app at app.adtest.ai.

What does an analysis cost through Claude?

The same as through the API: 1 credit per successful analysis, billed to the wallet tied to your API key. A failed analysis is never charged. The API and MCP interface are not included in the free trial.

Which clients can I use it with?

Claude Desktop and Claude Code out of the box, and any MCP client that can run a local stdio server with Node 18 or newer, such as Cursor or Windsurf.

Can it score video ads?

Yes. Video analysis runs as a background job and can take a few minutes. The server polls for the result and returns the report when it is ready, so long files are not cut off.

What can I pass to the analyze_advert tool?

Exactly one of: a public image or video URL, a local file path, or ad copy as text. You can add the advertiser's website as brand_url to improve brand and logo detection.

Is the server open source?

Yes. The code is MIT licensed at github.com/menaker/adtest-mcp, published on npm as adtest-mcp, and listed in the official MCP Registry as ai.adtest/adtest-mcp.

Score any ad from inside Claude