[BUG] MCP tools/call returns "Tool call failed: 500" against server that responds with HTTP 200

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Calling certain tools on a remote HTTP MCP server (https://quranx.com/mcp) from Claude Code returns Tool call failed: 500, even though the server itself responds with HTTP 200 and a valid JSON-RPC payload. The same calls succeed from direct curl, from a local MCP client, and from ChatGPT's MCP client — only Claude Code fails. Two specific tools fail consistently (get_verse_root_word_analysis, get_arabic_root_word_analysis); other tools on the same server work fine. The two failing tools are the ones that return the largest structured-content payloads (~33 KB and ~104 KB) and contain Arabic Unicode inside the structured content. Refreshing the MCP tool list does not fix it.

What Should Happen?

Claude Code should receive the same HTTP 200 / JSON-RPC structured-content response the server is actually returning, and surface it to the model as a normal tool result.

Error Messages/Logs

Claude Code surface:

Tool call failed: 500
Server-side (Honeycomb, ASP.NET Core + OTel instrumentation): zero HTTP 500s on /mcp/. All requests recorded as 200. So the "500" is being synthesized by Claude Code's MCP client; it does not correspond to any actual HTTP 500 on the wire.

Direct repro showing the server returns 200:

$ curl -isk -X POST https://quranx.com/mcp -H "Content-Type: application/json; charset=utf-8" -H "Accept: application/json, text/event-stream" --data-binary '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"get_verse_root_word_analysis","arguments":{"verseReference":{"chapter":4,"verse":34}}}}'
HTTP/1.1 200 OK
Content-Type: text/event-stream
...
event: message
data: {"result":{"content":[...],"structuredContent":{...}},"id":1,"jsonrpc":"2.0"}
(~33 KB body, ~150 ms.) The same payload returned for get_arabic_root_word_analysis with arabicRootWord: ضرب is ~104 KB. ChatGPT's MCP client receives and renders both responses without issue.

Steps to Reproduce

1. Add https://quranx.com/mcp as a remote HTTP MCP server in Claude Code.
2. Ask Claude to call get_verse_root_word_analysis for verse 4:34, e.g. "Use QuranX to get the word-by-word analysis of Quran 4:34."
3. Observe Tool call failed: 500 in the tool result.
4. Repeat with get_arabic_root_word_analysis for arabicRootWord: ضرب — also fails.
5. Compare with get_verses for the same verse — succeeds.
6. Run the curl above to confirm the server returns HTTP 200 with the full structured payload, and (optionally) make the same call from ChatGPT's MCP client to confirm it succeeds there too.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

1.5354.0

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_