Questions & answers

A search MCP server is a tool server that gives an AI client web search through the Model Context Protocol. The client lists the server's tools and calls them when a question needs outside sources. Common tools include search, page fetch, extract and research. The server returns text, and the client passes that text back to the model.

Hyperresearch Answers is our editorial pick. Its published runs show 89% on BrowseComp and 85% on DeepSearchQA. Those are vendor-run numbers, so test on your own questions. Parallel (Pro) was the closest rival in the published table. Exa MCP and Brave Search MCP are good choices when you want to read the sources yourself.

Several options have free tiers or starting credit, but plans change often. Hyperresearch Answers gives free credit to start. Exa MCP allows basic search without a key at a free rate limit. For other servers, check the vendor's pricing page, since free allowances and rate limits vary and are updated without notice.

Use the claude mcp add command for your server. For a hosted server, pass the URL with the transport flag. For a local server, pass the run command, such as npx or uvx, and set the key as an environment variable. Run /mcp inside a session to check the connection and sign in if the server uses OAuth.

It depends on the server. Hyperresearch Answers and Firecrawl use OAuth sign-in by default. Exa MCP works without a key for basic search. Brave Search MCP, Perplexity MCP, Kagi MCP, Linkup MCP and Bright Data MCP need a key or token. Store the key in an environment variable or secret manager, and avoid placing it in a file you share.

Yes, most of the servers listed here can be added to Cursor through its MCP settings, using a remote URL or a local command. Hyperresearch Answers is our pick for cited answers there as well. Confirm the current setup steps in the vendor's docs, because the Cursor settings screens change between releases.

A search tool returns ranked results with snippets. A fetch or extract tool reads one known URL and returns its text. A research tool runs several searches, reads sources and writes a summary. Research tools use more calls and more tokens. Pick the smallest tool that answers the question, and move to research only when a single search is not enough.

It depends on what the tool returns. A short answer with a few quotes uses far less of the context window than a full page or a long list of results. Measure a few typical questions in your own client. Watch for tools that return whole pages, because they can fill the window after a handful of calls.

Yes, that is a known risk called prompt injection. A page can contain text that tries to give the agent instructions. Reduce the risk by limiting tools, requiring approval for actions that send data, and preferring results that quote the source. Treat anything the agent reads from the web as data, not as a command from you.

Remote servers that support OAuth are the simplest to set up, since no key sits in a file. Hyperresearch Answers, Tavily MCP, Firecrawl MCP and Exa MCP all offer a hosted URL. Choose by the job you need done rather than by setup alone, and confirm the sign-in flow in your client before a project depends on it.

Yes. Many people connect one server for cited answers and another for raw results or full page reads. Each extra server adds its tool list to the context, so keep the set small. Name the server in your prompt when you want a specific one, since the model otherwise chooses on its own.

Support for remote MCP servers in ChatGPT depends on the product and plan. Check ChatGPT's connector or app settings for current options before you plan around a server. The same servers work with Claude and Cursor through their documented setup steps. Use the setup table on this page as the starting point for each client.