Give Cursor live web search, news and current docs with one MCP server
Cursor's agent writes code from training data that can be months old. Add one remote MCP server and it can look up current docs, read any page in full and check today's news before it answers. Setup takes one JSON block.
Cursor's agent is fast, but it writes code from what its model learned during training. Ask it about a library that shipped a breaking change last month and it will often give you the old API with total confidence.
The fix is to let the agent look things up. With the Model Context Protocol (MCP), you add tools to Cursor once and the agent decides when to call them. This guide adds one remote MCP server that gives Cursor:
- Web search from an independent index, for docs, guides and error messages.
- Page reading: the full readable text of any URL, so the agent reads the actual docs page instead of a snippet.
- News search by key words, date range and country, for anything that happened recently.
- Research papers from arXiv, PubMed and other open databases.
Step 1: get a key
The server used here is NestDaddy MCP. A free 7-day trial key works (5,000 calls, no card). Each tool call counts as one call; connecting and listing the tools is free.
Step 2: add the server to Cursor
Open ~/.cursor/mcp.json to make it available in every project, or create .cursor/mcp.json inside one project. Add:
{
"mcpServers": {
"nestdaddy": {
"url": "https://mcp.nestdaddy.com/mcp",
"headers": { "Authorization": "Bearer YOUR_KEY" }
}
}
}
If the file already has other servers, add the "nestdaddy" entry next to them inside mcpServers.
Then open Cursor Settings and find the MCP section (named "Tools & MCP" or "MCP" depending on your Cursor version). The server should show as enabled with a list of tools. If it does not, check the key and that the JSON is valid (a missing comma is the usual cause).
Keep the key out of git: if you use a project-level .cursor/mcp.json, add it to .gitignore.
Step 3: use it from the agent
MCP tools are available to Cursor's agent. You do not need to name the tool; describe what you need and the agent picks it. A few prompts that work well:
Read the current docs for the latest Next.js release and update this route handler
to the recommended pattern. Cite the docs page you used.
This error started after upgrading pydantic. Search for it, read the migration guide,
and fix the model definitions.
What changed in the GitHub Actions setup-python action this year?
Check the official docs before editing the workflow.
Find recent news about the Stripe API version we use and tell me if anything
affects our webhook handling.
The pattern that matters: ask the agent to read the source and cite it. That turns "the model remembers something" into "the model checked".
Step 4: tell the agent when to look things up
Agents call tools more reliably when the rules say when to use them. Add a project rule (for example .cursor/rules/research.mdc) with something like:
Before using a library API you are not sure about, search the web for the current
docs and read the relevant page in full. Prefer official documentation.
For anything that may have changed recently (releases, deprecations, prices,
outages), use news search with a recent date range.
Always list the URLs you relied on at the end of your answer.
Three sentences, and the agent stops guessing on version-sensitive code.
What the tools are good at
| You need | Ask for | Why it helps |
|---|---|---|
| Current docs for a library | web search + read page | The agent reads today's docs, not last year's |
| An error message fix | web search | Finds issues, discussions and changelogs |
| "Did X change recently?" | news search with a few days back | Recent releases and incidents by date |
| Background on an algorithm | research papers | Primary sources instead of blog summaries |
Cursor's built-in web search, or an MCP server?
Cursor can also search the web on its own, and for quick lookups that is often enough. An MCP server is worth adding when you want:
- The same tools everywhere: the same server works in Claude, Claude Code, VS Code, Windsurf and n8n with the same key.
- News with filters: key words plus a date window and a country.
- Full page text from any URL, and research papers.
- Control and logs: you see each tool call and what it returned.
They do not exclude each other. Many people keep both.
Troubleshooting
- Server shows an error or no tools: the key is wrong or expired, or the JSON has a syntax error. Paste the file into a JSON validator.
- The agent never calls the tools: add the project rule from Step 4, or ask explicitly ("search the docs first").
- "Quota exceeded": the monthly call limit was reached. Calls return that message instead of failing silently.
- Corporate proxy: the server is plain HTTPS on port 443 at
mcp.nestdaddy.com.
Setup for other clients
The same server works in Claude Desktop, Claude Code, VS Code, Windsurf and n8n. Copy-paste setup for each one is on the NestDaddy MCP page, and the full walkthrough for Claude is in How to give Claude live web search and news with one MCP server.
Plans start at $9.99 a month for 10,000 calls after the free 7-day trial.
