How to Set Up MCP in Cursor

MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. is an open standard for connecting AI agents to external tools and data sources. In Cursor, an MCP server exposes tools (query a database, read a ticket, call an API) that the agent can use mid-task - so it isn't limited to your codebase.

The practical framing from people who use this daily: MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers are how you connect Cursor to your internal knowledge bases. Think in two buckets. Personal MCPs are tools only you use (Figma, a meeting-notes service like Granola). Team MCPs are the shared systems your whole org reads from - your data warehouse (e.g. Databricks), your experiment platform (e.g. Statsig), your backend services. Once an MCP is connected you can even steer a confused agent back on track mid-run by telling it which one to use, like "use the Databricks MCP."

Connecting a server is not free from the agent's side, and the two-bucket framing leaves that out. Every enabled tool is another option the model weighs on each turn, and Cursor's own advice for an agent that keeps reaching for the wrong thing is to trim the enabled set back to what you need. Keep the personal bucket small, and switch team servers on for the work that uses them instead of leaving all of them lit.

You point Cursor at the server, give it any keys it needs, and confirm its tools show up. Add it in Settings → Tools & MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. or write a .cursor/mcp.json in your project; a global ~/.cursor/mcp.json covers every project, and the project file wins when names clash. A local server gets a command to launch; a hosted one gets a URL.

1. 1Open Cursor Settings → Tools & MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. (or create a .cursor/mcp.json in your project).
2. 2Add the server: a command (stdio, e.g. npx -y some-mcp-server) or a remote URL (streamable HTTP or SSE).
3. 3Provide any required env vars / API keys - marketplace remote servers can authenticate with OAuth instead.
4. 4Reload; toggle the server on and confirm its tools show as enabled.
5. 5Ask the agent to use a tool - by default it requests approval before calling it.

{
  "mcpServers": {
    "example": {
      "command": "npx",
      "args": ["-y", "@example/mcp-server"],
      "env": { "API_KEY": "..." }
    }
  }
}

Which of the two files you write it into deserves a moment of thought. A global ~/.cursor/mcp.json follows you into every project, and the project file wins when both define the same server name, so a global entry can sit there being shadowed while you edit it and wonder why nothing changes. The project file also sits inside the repository, so an API key pasted into env is one commit away from travelling with every clone. Marketplace remote servers authenticate over OAuth instead and sidestep that question.

The most useful MCPs are the systems your team already lives in. A few that show up again and again, with the workflow each one enables:

To see what's available, browse the Cursor Marketplace for official plugins first, then cursor.directory for community servers - between them, that's the discovery surface for checking whether a given tool (Excalidraw, Lucidchart, whatever) has a server.

The Figma MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. runs both directions, which is why designers reach for it. Inbound, you paste a Figma file link into the prompt - "use the design system at this link to inform the design of my dashboard" - and Cursor reads the tokens, typography, colors, and variable definitions, takes screenshots, and applies them to your live code. Outbound, you select a rendered component in design mode, ⌘+L it into chat, and say "publish this component back to Figma" with the target file link. Cursor recreates it there with the same colors, fonts, and width, split into proper layers; the published-back layers show up in purple.

Almost always it means the server never started. The usual culprits are a bad command or path, a missing API key, or a stdio/remote mismatch. Run the exact command in a terminal first to see it start before you blame Cursor.

• The command/path is wrong or the package isn't installed - test it in a terminal first.
• Missing env vars / API keys so the server fails to start.
• stdio vs remote mismatch - a remote server needs a URL, not a command.
• Too many tools enabled across servers - trim to what you need so the agent isn't overwhelmed.
• Still stuck? Open the Output panel (⌘/Ctrl+⇧+U) and pick MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. Logs - the server's startup errors land there.