MCP Server Crashes or Fails on Large Responses: How to Fix It (Claude Code, Claude Desktop, Cursor)

When an MCP tool returns a very large result, three things can go wrong: the client caps or moves the output, a transport limit cuts the response off, or the server itself runs out of memory building it. The fastest fix is to make the tool return less per call, with filters, limits or a cursor. As a user of Claude Code you can also raise MAX_MCP_OUTPUT_TOKENS, but the lasting fix sits in the server.

Checked against: Claude Code's MCP output limits, the 2026-07-28 MCP spec pages on tools, pagination and the stdio transport, plus the MCP debugging guide, on 2 November 2026. eSEOspace did not run these steps.

What this error means

Large results hit limits at several layers. The model has a finite context window. The client protects it with token caps. The transport has message size limits. And the server has to hold the full result in memory before it sends it.

You might see a warning about output size, a note that the result was saved to a file, a "connection closed" error, a request timeout, or a server that exits mid-call. On stdio, each message is a single line of JSON. The stdio spec says messages "MUST NOT contain embedded newlines," so a server that pretty-prints a huge object can also break the stream.

Quick fix

1. Confirm the cause in the log: size warning, timeout, out-of-memory, or connection closed.
2. Ask for less. Narrow the date range, add a filter or request specific fields.
3. If you use Claude Code and need more room, start it with a higher limit: export MAX_MCP_OUTPUT_TOKENS=50000 then claude.
4. If you maintain the server, add limit and cursor arguments to heavy tools and return one page at a time.
5. Return a link to the full data instead of the data itself, using a resource_link content item.
6. Restart the client and repeat the call.

Fix by client / OS

Claude Code

Claude Code's docs set out the limits in detail:

A server can raise the threshold for one tool by adding _meta to its tools/list entry:

{
  "name": "get_schema",
  "description": "Returns the full database schema",
  "_meta": {
    "anthropic/maxResultSizeChars": 200000
  }
}

The docs say images are still subject to MAX_MCP_OUTPUT_TOKENS. They also describe a startup timeout (MCP_TIMEOUT) and a per-server tool timeout field in .mcp.json, which help when a big result is slow rather than oversized.

Claude Desktop

Claude Desktop does not publish output limits in the pages we checked. Read mcp-server-SERVERNAME.log for the server's stderr. If the server process died, the log usually shows its last error. Reduce the result size on the server side, as below. The config file is ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows.

Cursor and VS Code

Neither Cursor's MCP docs nor the VS Code MCP pages we opened list a tool output size limit. Use their logs to see whether the server crashed or the call timed out, then apply the server fixes below.

Server fixes (any client)

• Paginate inside your tool. The pagination spec covers resources/list, resources/templates/list, prompts/list and tools/list. For tool results, add your own cursor argument and return a next cursor in the result. Our MCP pagination guide explains the model.
• Return links, not blobs. The tools spec lets a tool return a resource_link with a URI the client can fetch later. See MCP content types.
• Return structured, trimmed data. Drop unused fields. structuredContent with an outputSchema makes the shape predictable.
• Stream to disk, not memory. Build big exports as files and return their location.
• Never pretty-print over stdio. Serialize each message on one line.
• Log to stderr only. The MCP debugging guide says local servers should not log to stdout.

Where to find the logs

• Claude Code: /mcp shows server status. Results saved to disk go to the session's tool-results directory under ~/.claude/projects/.
• Claude Desktop: ~/Library/Logs/Claude (macOS), %APPDATA%\Claude\logs (Windows), ~/.config/Claude/logs (Linux).
• Cursor: Output panel, then MCP Logs.
• VS Code: MCP: List Servers, select the server, Show Output.

If it still fails

• Timeouts instead of size errors. Slow queries look like crashes. See MCP request timed out.
• Connection closed mid-call. Usually the process exited. See MCP error -32000 connection closed.
• The tool list itself is huge. Too many tools use context before any call. See too many tools warning.
• Resource limits as a security issue. Unbounded responses can be abused. Our page on MCP resource exhaustion covers safe limits.

Look up any JSON-RPC error code you see with the MCP error code lookup.

Next step

If you run an internal MCP server that returns reports, schemas or exports, design every heavy tool with a limit and a cursor from day one. eSEOspace's MCP development team builds servers this way and can review an existing one.

Frequently asked questions