MCP error -32000: Connection closed: How to Fix It (Claude Desktop, Cursor, Claude Code)

MCP error -32000: Connection closed means the client was waiting on your MCP server and the connection ended first. For a local stdio server, that almost always means the server process exited, either at startup or mid-request. The fastest fix is to run the exact command and args from your config in a terminal and read what it prints before it quits. Then check the server's own log file.

Checked against: the source of the @modelcontextprotocol/sdk 1.32.2 npm package, the MCP 2026-07-28 base protocol page, the stdio transport spec, the MCP debugging guide and the Claude Code MCP docs on 2 November 2026. eSEOspace did not reproduce this error in each client.

What this error means

The JSON-RPC 2.0 spec reserves -32000 to -32099 for "implementation-defined server-errors." MCP's TypeScript SDK picked -32000 for a closed connection. In the 1.32.2 package, the error class formats messages as MCP error <code>: <message>, which is exactly the text you see.

The SDK throws it for every request still waiting when the transport closes. It also uses the same code for "Request was cancelled" in some paths. So the code alone tells you the connection is gone, not why.

The 2026-07-28 base protocol page adds an important caveat. It calls -32000 to -32019 "legacy" and says receivers "MUST NOT assume any specific meaning" for those codes. A remote server might send -32000 for an unrelated server error. Always read the message text with the code.

For stdio servers, the transport spec says that if the server process exits unexpectedly, the client "SHOULD restart it," and in-flight requests "are simply lost."

Quick fix

1. Run the server by hand. Copy command and args from your config into a terminal. A missing module, bad flag or Python traceback appears right away.
2. Read the server's stderr log. In Claude Desktop that is mcp-server-SERVERNAME.log.
3. Check required environment variables. A server that needs an API key often exits when the key is missing. Add it under env.
4. Remove stdout output. Any print() or console.log() in a stdio server breaks the stream. See stdout logging breaks the server.
5. Use absolute paths. The debugging guide warns that the working directory may be / on macOS.
6. Restart the client fully and try again.

Fix by client / OS

Claude Desktop

Logs live in ~/Library/Logs/Claude (macOS), %APPDATA%\Claude\logs (Windows) and ~/.config/Claude/logs (Linux). Look in mcp.log for the time the connection closed, then open the server's own log for the same minute. If the server ran under the app's built-in Node runtime as a desktop extension, a bug report on SAP's Fiori MCP server shows a case where pointing command at a system Node fixed a silent exit. Treat that as one user's report.

Cursor

Cursor's MCP docs say Cursor "isolates server failures," so other servers keep working. Open Output, choose MCP Logs, and look for "connection errors, authentication issues, or server crashes." Toggle the server off and on under Customize to restart it.

Claude Code

The Claude Code MCP docs say a stdio server whose first connection fails is marked as failed with no automatic retry. Select Reconnect in /mcp after fixing it. For servers that drop later, Claude Code retries, and after five failed attempts shows MCP server "<name>" disconnected · open /mcp to reconnect. claude mcp get <name> prints the error on an Issue: line.

VS Code

Run MCP: List Servers, pick the server and choose Show Output, as the VS Code MCP management page describes. Use Restart Server after a fix.

Where to find the logs

If it still fails

• It closes after exactly one minute. That is probably a timeout, not a crash. The SDK's default request timeout is 60,000 ms. See MCP request timed out.
• It closes only on large results. The 1.32.2 SDK's stdio reader drops and closes when its buffer passes 10 MB. Return smaller results or paginate.
• It works in the Inspector. Test in the MCP Inspector. If it passes there, read works in Inspector but not in Claude.
• It is a remote server. A closed HTTP connection can be a proxy timeout or a crashed host. Check host logs and the streamable HTTP transport rules.

Decode any other code with the MCP error code lookup. If you build servers for clients and want them to fail cleanly, see MCP error handling or talk to us about MCP development.

Frequently asked questions