MCP server

Google Trends inside your assistant. TrendFlow speaks the Model Context Protocol over Streamable HTTP, so any MCP client can use it without installing anything.

Endpoint

https://trendflow.mory.dev/mcp

Connecting

In Claude Desktop or Cursor, add it as a remote MCP server. In ChatGPT, add it as a connector. In Perplexity it is a custom remote connector, filled in through a form rather than a config file — see Google Trends in Perplexity.

{
  "mcpServers": {
    "trendflow": { "url": "https://trendflow.mory.dev/mcp" }
  }
}

Tools

Free and paid

Topic lookup and the trending RSS feed are free and need no account — they cost nothing to serve, because neither needs a proxy. The tools that query Google's historical endpoints do, so those need a connected TrendFlow account. The first time your assistant reaches for one, it answers with a link to connect rather than an error.

Connecting an account

Go to trendflow.mory.dev/connect. Signing in creates a free account if you do not have one, and hands back a configuration ready to paste.

The configuration adds an API key as a bearer token, which every MCP client supports:

{
  "mcpServers": {
    "trendflow": {
      "url": "https://trendflow.mory.dev/mcp",
      "headers": { "Authorization": "Bearer tfk_your_key" }
    }
  }
}

In opencode the block is named mcp rather than mcpServers and the server needs "type": "remote":

{
  "mcp": {
    "trendflow": {
      "type": "remote",
      "url": "https://trendflow.mory.dev/mcp",
      "headers": { "Authorization": "Bearer {env:TRENDFLOW_API_KEY}" }
    }
  }
}

Reading the key from the environment, as above, keeps it out of a file you might commit.

Clients that implement the MCP OAuth flow can instead sign in through the browser — the server advertises its authorization server at /.well-known/oauth-protected-resource and supports dynamic client registration with PKCE.

Narrowing a query

The three historical tools take four optional filters beyond the keyword. They compose, so you can ask for YouTube interest in one US state over a custom range of dates.

The properties are separate indexes, not filters over one dataset, so a term can be quiet on web search and busy on YouTube. Values from two different properties are not comparable to each other. The timeframe also sets granularity — an hourly range returns minute-level points and all returns monthly ones, so a long range cannot show a short spike.

Reading the numbers

Every figure is normalized relative interest, scaled 0–100 within a single result set — never absolute search volume. Results from two separate queries are not comparable to each other.

Freshness

Responses are cached, from 15 minutes for a one-day series up to 12 hours for multi-year history, which barely moves day to day. If Google rate-limits us we serve a slightly older answer rather than an error.

Self-hosting

The MCP server is open source. npx trendflow-mcp speaks stdio, and npx trendflow-mcp --http=8080 serves the same tools over Streamable HTTP. You bring your own proxies — see Proxies and rate limits.

Questions people ask