ChatGPT reaches the Trends API MCP server through its connector flow. Add an app with the URL https://trendsapi.ai/mcp, sign in on Trends API to authorize, and the three tools appear in chat: get_time_series for keyword history, get_growth for percent change over preset windows, and get_top_trends for live boards. The authorized key is the same product key as the REST API. Server-level detail sits on MCP server setup.

Add the app in ChatGPT

Open the app or connector settings in ChatGPT and create a new connection. Name: Trends API. MCP server URL: https://trendsapi.ai/mcp. The quickstart on the homepage opens that form with the fields filled in. Confirm, then complete the sign-in on Trends API. After authorization the tools list shows the three modes. No config file is edited in this flow, and nothing is installed locally.

Ask in source words, not parameter names

The model translates a plain question into tool arguments. "How much did searches for heat pumps grow over 12 months" becomes a get_growth call with source google search. "What is trending on TikTok right now" becomes get_top_trends with type TikTok Trending Hashtags.

Two vocabularies matter. Keyword history takes a source string such as google search, tiktok, or amazon. Live boards take a type label such as Google Trends or YouTube Trending. The strings are not interchangeable, and the full list sits on the sources reference. Naming the platform in the question removes the guess.

What the answers are made of

A history answer is weekly points: a date, a 0-100 value, and volume when the source publishes it. A growth answer is one percent per requested window, against the window start. A board answer is a ranked list as of a timestamp. Asking for two sources in one question produces two calls and one combined reply, which is the practical reason to run trend questions in chat instead of a notebook.

When REST is still the right tool

Schedulers, Airflow dags, and backfills should call POST https://api.trendsapi.ai/api directly. A chat session adds a model in front of a request a script can already send, and the reply is prose, not a stored table. Only successful 200 responses count against quota; caps sit on pricing. The Claude version of this setup sits on Claude Trends MCP.