How to Add an MCP Server to Cursor

Cursor 0.40+ supports Model Context Protocol (MCP) servers natively, you can plug in structured data tools that the AI inside Cursor can call during conversations.

Step 1, Open the MCP settings. Click the Cursor settings icon (bottom-left or Cmd+, on macOS), navigate to Tools → MCP. You'll see an editable JSON config area.

Step 2, Paste the server config. For the GitDealFlow MCP server (@gitdealflow/mcp-signal on npm), add an entry whose command runs npx -y @gitdealflow/mcp-signal. The full six-tool server, trending startups, sector search, single-startup signal, dataset summary, scout receipts, and methodology, starts with no API key and no auth of any kind.

Step 3, Save and restart. Cursor re-reads the MCP config on restart; once the server shows green, the tools are callable from any chat or agent session.

Step 4, Try a sourcing query. Ask Cursor something like "using the gitdealflow tools, list trending AI startups this week and pull the signal for the top three." The agent resolves the intent to get_trending_startups and get_startup_signal calls, and you get structured JSON you can drop straight into a memo or a sheet.

What this unlocks for deal sourcing. The point of MCP in Cursor is not novelty, it is removing the copy-paste layer between a dataset and your working context. Instead of toggling to a browser, exporting CSV, and re-formatting, the numbers you would have looked up are already inside the conversation where you are writing the investment thesis. For a weekly sourcing ritual, that collapses a ten-minute context switch into a one-sentence request.

Verification and troubleshooting. If the server shows red, run npx -y @gitdealflow/mcp-signal once in a terminal so the package caches, then restart Cursor; first-run downloads are the most common failure. If tools are visible but return errors, confirm network access to signals.gitdealflow.com, the server reads the public JSON endpoint directly. The server is read-only and idempotent, so there is no state to corrupt and nothing to undo: any call can simply be re-issued.

The config is portable across hosts. The same mcpServers JSON shape works in every MCP-compatible host, so the entry you paste into Cursor can be copied into Claude Code's .claude/mcp.json project file with no changes, or into Claude Desktop's config on macOS or Windows. This matters because most investors run more than one AI host, and the dataset should follow you rather than being re-configured from scratch each time. The GitDealFlow server is also listed in the official Model Context Protocol Registry under the canonical identifier io.github.kindrat86/vc-deal-flow-signal, so the package and the registry entry can be cross-checked.

Multiple servers coexist cleanly. Cursor supports any number of entries inside the single mcpServers object. Each server runs as its own process, and the AI host decides which tools to call based on the question it is answering. A typical sourcing setup pairs the GitDealFlow server with a web search server and a GitHub server, giving the agent raw repository access alongside the structured signal data, with the tools appearing side by side in the same toolbox.

Network and caching behavior. The MCP server fetches live data from the public dataset endpoint, so it requires internet access and will not work fully offline; the npm package itself can be cached locally after the first install. The most common first-run failure is a missing cached package, which is why running npx -y @gitdealflow/mcp-signal once in a terminal before restarting Cursor resolves most red-state errors. If tools appear but return errors, the next check is network reachability to signals.gitdealflow.com, where the server reads its public JSON endpoint.

Removal is symmetric to installation. To remove a server, delete its entry from Settings, Tools, MCP, and restart Cursor; the tools disappear from the agent toolbox on the next session. If you want to fully clean the package, npm uninstall -g @gitdealflow/mcp-signal clears the global npm cache. Because the server is read-only and requires no API key, uninstalling leaves no credentials or state behind, which keeps the security surface trivial.