Cursor's agent mode becomes significantly more capable when it can call MCP tools directly during a task. Instead of asking you to run a command or fetch data manually, the agent fetches it inline, uses the result, and continues. This guide shows the exact config format and which servers are worth connecting first.
For background on what MCP is and how the protocol works, read What Is the Model Context Protocol. The devtools category alone covers 2,349 servers on MCPFind - this guide helps you cut through that and start with what matters. Once you have Cursor configured, the same patterns apply in VS Code. If you are coming from Claude Desktop, getting started with MCP in Claude explains the same install flow for that client.
Where Cursor Stores MCP Config
Cursor reads MCP server config from a JSON file. You can place it at the project level or globally.
Project-level config lives at .cursor/mcp.json inside your repository root. This is the right choice for project-specific servers like a database server tied to one app. Global config lives at ~/.cursor/mcp.json and applies across every project you open in Cursor. Most developers use global config for general-purpose servers (search, filesystem, documentation) and project-level for anything tied to a specific stack.
The JSON Config Format
The config format is straightforward. Here is the structure for a single server:
{
"mcpServers": {
"server-name": {
"command": "npx",
"args": ["-y", "@scope/package-name"],
"env": {
"API_KEY": "your-key-here"
}
}
}
}The command field is the executable Cursor will run. npx is the standard for Node-based servers. Python servers typically use uvx or python. The args array passes arguments to the command. The env field is optional - use it for API keys and secrets so they stay out of your args.
Here is a real example using the Supabase MCP server, which has 2,556 GitHub stars and is the top-rated server in /categories/databases:
{
"mcpServers": {
"supabase": {
"command": "npx",
"args": ["-y", "@supabase/mcp-server-supabase@latest"],
"env": {
"SUPABASE_URL": "https://your-project.supabase.co",
"SUPABASE_SERVICE_ROLE_KEY": "your-service-role-key"
}
}
}
}How to Configure MCP in Cursor: Which Servers Work Best
We looked at the 2,349 servers in /categories/devtools ranked by GitHub stars. Three categories of servers add the most value in Cursor's agent mode.
The first is diagram and whiteboard tools. tldraw leads this group with 46,031 GitHub stars - the most starred server in the entire MCPFind index. It lets Cursor generate architectural diagrams inline during planning tasks. The second is browser and testing tools, where the Chrome DevTools MCP server (31,292 stars) enables Cursor to inspect live browser state during debugging sessions. The third is documentation search, where servers that query your internal docs or public API references let the agent look up correct method signatures without asking you.
For most Cursor setups, start with one database server and one search server. Add devtools-specific servers as you identify gaps in your workflow.
Verifying the Config Works
After saving your config file, reload Cursor with Cmd+Shift+P and select "Developer: Reload Window." Then open a new composer session and type a prompt that requires the server you just added.
If the server is connected, you will see tool call entries appear in the composer output as the agent runs. If nothing appears, open the Cursor output panel (View > Output) and select "MCP" from the dropdown to see connection errors. The most common issues are a missing npx in PATH (fix by using the full path like /usr/local/bin/npx) and malformed JSON in your config file. Validate your JSON with a linter before reloading.
Gus Marquez