For AI agents: the full docs index is at /docs/llms.txt. Append .md to any docs URL for Markdown.
Skip to content
Proxy Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Ads MCP for Claude: connector setup and tool loading

Set up Proxy's ads MCP in Claude on web, desktop and mobile: custom connector steps, how Claude loads the tools, permissions, sign-in and common fixes.

This page covers Claude on the web, the desktop app and the mobile app. For the terminal, see Ads MCP for Claude Code.

One connector gives Claude all four ad libraries: Meta, Google, LinkedIn and TikTok (EU).

Connect

  1. Copy the Proxy MCP URL:

    https://mcp.useproxy.dev/mcp?ref=docs
  2. In Claude, open Customize, then Connectors, then Add custom connector. This link opens the dialog: claude.ai/customize/connectors.

  3. Name it Proxy, paste the URL and sign in.

  4. When Claude asks for permission to use a Proxy tool, choose Always allow.

How sign-in works

  • There is no API key field to fill. Claude registers itself with Proxy and runs OAuth with PKCE.
  • The first connection opens a Proxy sign-in page with Google or email. A new email creates a free account in the same step.
  • The browser then returns to claude.ai, and Claude stores the token.
  • An unauthenticated request getting a 401 first is the normal handshake, not an outage. See Authentication.

How Claude loads the tools

Claude does not read every connector tool up front. Until it searches for a tool, it sees only tool names and the server’s short instructions. This matters in two ways.

Five entry tools are loaded from the start. Proxy marks them so Claude can call them without searching first:

  • get_meta_platform_id
  • search_meta_ads_by_keyword
  • get_google_platform_id
  • search_linkedin_ads
  • search_tiktok_ads

The rest, such as get_meta_ads, get_google_ads and the analysis tools, load when Claude needs them.

Claude can still pick web search. Proxy’s instructions tell Claude to use these tools, not web search, for any question about a brand’s ads. Web search finds articles about ads; Proxy returns the ads. If Claude answers from the web anyway, start your prompt with “Using Proxy”.

Permissions

Every Proxy tool is declared read-only and non-destructive. Proxy reads public ad libraries and never connects to your ad accounts. Always allow is safe to choose per tool, and it stops the prompt from repeating.

What you see in the chat

  • Ad carousels. Claude can render the ads it found as an inline carousel, up to 20 ads each, next to the point it is making.
  • Links. Each ad comes with a link back to the public library.
  • Same data everywhere. Results carry the same identifiers in the text and structured parts of the response, so the web, desktop and mobile apps all let Claude chain on them.
  • Shareable pages. Ask for a page to share with a client or team. Claude calls get_page_starter, which is free, and builds an artifact that reads live ads through Proxy.
  • Weekly tracking. Re-running a Meta or Google pull for the same brand reports what changed since your last fetch. Ask Claude to schedule it weekly.

Troubleshooting

Symptom Fix
Claude answers from web search Start the prompt with “Using Proxy” and name the platform
Claude says a tool is missing Ask it to load the Proxy tools. Only the five entry tools are loaded up front
A permission prompt on every call Choose Always allow for that tool
A removed tool is called, such as analyze_ad_videos_batch Proxy returns an error pointing at analyze_ad_video. Ask for the videos one at a time
Sign-in fails with a redirect error Email [email protected] with the redirect URI shown
Credits run out The tool returns what it gathered plus an upgrade link. Nothing is billed. See Credits and plans

Example prompts

Using Proxy, what Meta and Google ads is [brand] running right now? Group them by angle.

get_meta_ads, get_google_platform_id, get_google_ads.

Using Proxy, find five competitors of [domain] and pull their LinkedIn ads.

search_competitors, then search_linkedin_ads per company.

Turn that into a page I can share with my client.

get_page_starter, then Claude builds the artifact.

Guides

Was this page helpful?