A hosted MCP server for AI answers and AI Overviews

#ai #api #llm #mcp

AI agents are increasingly the consumer of search data. If your tooling already speaks MCP — Claude Desktop, Cursor, an agent framework — wiring it to a hosted MCP server is less code than a REST client.

This post shows a hosted MCP server that returns what ChatGPT, Gemini, Google AI Mode and Google Search (including AI Overviews and News) answer, as structured JSON with sources. I work on it — it's called AnswerLine.

The endpoint

https://mcp.answerline.dev/mcp        (streamable HTTP)
Authorization: Bearer <key>

For MCP clients that can't set headers, the key also works in the path: https://mcp.answerline.dev/<key>/mcp.

The server is listed in the GitHub repo and exposes tools, not resources — every call is one structured request:

What the tools return

Tool	Surface	Returns
`chatgpt`	ChatGPT	answer text, cited sources, ads, shopping cards, search queries
`gemini`	Google Gemini	answer text, cited sources, markdown export
`google`	Google Search	organic listings, PAA, related searches, optional AI Overview
`aimode`	Google AI Mode	answer text, sources, product clusters
`google_news`	Google News	title, link, snippet, source, date per article
`credits`	account	remaining credits and billing cycle

A chatgpt call asks for {prompt, country} and comes back with the answer plus everything the UI showed: which sites were cited, which queries the model fanned out to, which ads and shopping cards appeared next to the answer. That last part is what makes it useful for visibility tracking — the citation list is the data you can't get from a SERP API.

It's plain JSON-RPC over HTTP

Any MCP client that supports streamable-HTTP remote servers can point at the endpoint. Underneath it's standard JSON-RPC — you can even drive it with curl:

curl -X POST https://mcp.answerline.dev/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'Authorization: Bearer YOUR_KEY' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

tools/list and initialize are readable anonymously, so MCP registries can index the server without a key.

A weekly brand-visibility check

The pattern I actually use: a scheduled job asks every surface the same buyer-intent prompt set, diffs the citation lists week over week, and alerts when a competitor gets cited where you used to be. The public version of this runs weekly across ChatGPT, Gemini and AI Mode and publishes the result on a live leaderboard — in half the tracked categories the engines disagree about who the leading brand is.

Free tier

500 one-time credits, no card — sign up at the dashboard. Failed requests aren't charged; pricing is per call on the pricing page.

DEV Community