Trends MCP for ChatGPT

Trends MCP is the data layer. ChatGPT is a conversation host that can attach it as a custom connector. That connector should use https://api.trendsmcp.ai/mcp with Authorization: Bearer YOUR_API_KEY. ChatGPT Settings → Connectors → Add custom connector is the UI. Once the tools appear, ChatGPT can call get_trends (weekly series), get_growth (percent windows), and get_top_trends (live feeds). The free Trends MCP plan is 100 requests per month. Starter is $19 per month. ChatGPT is the right client when the work is a long back-and-forth research chat with no repository open, and the failure mode to avoid is a fluent answer that stopped being true at the model’s training cutoff.

Connector setup without mixing in Claude’s OAuth host

Add a custom connector named Trends MCP. Server URL: https://api.trendsmcp.ai/mcp. Supply the Bearer API key the same way other remote MCP connectors in ChatGPT take secrets. Replace YOUR_API_KEY with the tmcp_liv… token from account signup. Enable the connector for the chat that needs it.

Claude.ai is documented separately because it is easy to “correct” the URL. Claude.ai Connectors use OAuth against https://www.trendsmcp.ai/mcp. That www host is intentional for Claude.ai and wrong for ChatGPT. ChatGPT stays on the api host with a key. Claude Desktop is a third configuration (npx mcp-remote plus AUTH_HEADER). None of those Desktop JSON files belong in ChatGPT. The Connect docs keep the three families apart.

If the connector saves but tools never run, confirm the workspace allowlist includes api.trendsmcp.ai, confirm Plus or a higher tier actually exposes custom connectors in that account, and confirm the prompt says “using TrendsMCP.” ChatGPT otherwise answers from memory and never touches the server.

When ChatGPT is the right host

ChatGPT fits operators who already think in threads: market maps, content calendars, investor FAQs, product naming. They want the model to argue, rewrite, and compare, with numbers that did not come from a 2024 snapshot. An IDE is overhead. Raycast is too small. n8n is too much graph.

Cursor and VS Code remain better when the next token should be code. Perplexity remains better when the user wants the model to mix web pages with numbers in a citation-heavy research product. ChatGPT remains better when the user already lives in ChatGPT projects, custom GPTs, or Team workspaces and only needs the trend server as one connector among others.

Trends MCP does not make ChatGPT a licensed Google Trends console. It makes ChatGPT able to call the same three tools every other MCP client calls.

How a research thread should call the tools

Start with a live feed if the topic is unknown: get_top_trends with a precise type such as Google Trends or TikTok Trending Hashtags. Then get_growth on two or three surviving phrases, with windows like 3M and 1Y in one call. Finish with get_trends only for the phrase that will be quoted as a curve. REST automations that mirror this should send mode: "get_time_series" for the history step.

Ask ChatGPT to print the tool arguments before calling them. The model likes source: "Google Trends" on a growth call. That string is a feed type, not a keyword source. Keyword sources include google search, youtube, tiktok, reddit, amazon, npm. Feed types are a different table on data sources.

A ChatGPT Project can store a standing instruction: always use TrendsMCP for momentum questions; never invent a source label; budget eight calls per thread. That instruction matters more here than in Cursor because ChatGPT users ask broader questions.

Avoid source='all' in casual threads. Each source plus keyword is a request. A generous “compare every platform” prompt can spend the month in one sitting.

Plans, admins, and double billing

ChatGPT Plus, Team, or Enterprise is OpenAI’s seat. Trends MCP Free, Starter ($19 per month), Pro, or Business is a second bill. Enterprise ChatGPT admins may block custom MCP URLs. The connector will look fine to an individual and still never fire. Get https://api.trendsmcp.ai/mcp allowlisted.

Do not paste API keys into the chat. Keys belong in the connector secret field. If a key leaked in a shared Team transcript, rotate it on the Trends MCP account page.

Limits of a conversational caller

The free 100 Trends MCP requests last a month. ChatGPT will retry tools. Failed 400s still often count if the call reached the API. Keep limit modest on live feeds. History is weekly over about five years, which is enough to argue seasonality and the wrong grain for “this hour.”

ChatGPT may summarize tool JSON incorrectly. Require the model to quote growth, recent_date, and baseline_date from get_growth before writing a narrative. A 900 percent claim with a baseline near zero should stay in the thread as a caveat.

Custom GPTs that embed the connector still share the same monthly quota as the account. Ten GPTs are not ten pools.

Temporary chats, shared links, and Team workspaces all still spend the same Trends MCP meter. A viral shared prompt that says “check every TikTok feed” will bill the connector owner, not the reader. Disable the connector on chats that are meant to be forwarded. Custom GPTs that embed the connector should document the expected call budget in the GPT instructions so testers do not treat the GPT as an unlimited dashboard.

ChatGPT is a strong conversation host for the shared Trends MCP server. It is not Claude.ai OAuth, not an IDE, and not a substitute for scheduled REST jobs.