[BUG] stdio MCP server always times out at ~30000ms on Windows, even though it connects via MCP Inspector)

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?

Bug Report: stdio MCP server always times out at ~30000ms on Windows, even though the server itself connects successfully via MCP Inspector

Environment

• Claude Code: v2.1.220 (native install)
• Install path: C:\Users\<user>\.local\bin\claude.exe
• Config install method: native
• Platform: win32-x64 (Windows), tested in standalone PowerShell 7.6.4 (Windows Terminal), not VS Code integrated terminal
• MCP server: Oracle SQLcl (sql.exe -mcp), stdio transport

.mcp.json

\\\json
{
"mcpServers": {
"sqlcl": {
"type": "stdio",
"command": "C:\\sqlcl-latest\\sqlcl\\bin\\sql.exe",
"args": ["-mcp"]
}
}
}
\\\

What's Wrong?

The "sqlcl" stdio MCP server always fails to connect from Claude Code CLI with a generic timeout, even though the exact same binary/command connects successfully and quickly (~11s) when tested through the official MCP Inspector (npx @modelcontextprotocol/inspector <command>).

Debug log (relevant lines)

\\\
[DEBUG] MCP server "sqlcl": Starting connection with timeout of 30000ms
[DEBUG] MCP server "sqlcl": Connection timeout triggered after ~30000ms (limit: 30000ms)
[DEBUG] MCP server "sqlcl": Connection failed after ~30000ms: MCP server "sqlcl" connection timed out after 30000ms
[ERROR] MCP server "sqlcl" Connection failed: MCP server "sqlcl" connection timed out after 30000ms
\\\
No additional stderr/child-process output is surfaced in the debug log for this stdio server (unlike HTTP-transport servers, which log Environment/transport details — see e.g. the "github" server's HTTP transport logs in the same session).

Steps to Reproduce

1. Configure a native .exe as a stdio MCP server in .mcp.json (see above).
2. Start Claude Code in the project directory: claude
3. Run /mcp — server shows ✘ failed after ~30s.

What I've already ruled out

• Server-side issue: Manually running C:\sqlcl-latest\sqlcl\bin\sql.exe -mcp in a terminal starts instantly and prints "MCP Server started successfully".
• stdout pollution: Redirected stdout/stderr separately (sql.exe -mcp > out.txt 2> err.txt); confirmed the startup banner goes to stderr only — stdout is clean.
• MCP protocol implementation issue: Verified with npx @modelcontextprotocol/inspector "C:\sqlcl-latest\sqlcl\bin\sql.exe" -mcp — connects successfully in 11305ms, handshake completes, capability negotiation (MCP 2024-11-05) succeeds.
• Resource contention from multiple stdio servers: Isolated test with only "sqlcl" enabled (removed "playwright" and "desktop-commander" via claude mcp remove ... -s local) — still times out identically.
• env key stripping parent environment: Removed "env": {} from .mcp.json entirely — still times out identically.
• Shell wrapping: Changed command to "cmd" with args: ["/c", "C:\\sqlcl-latest\\sqlcl\\bin\\sql.exe", "-mcp"] — still times out identically.
• VS Code integrated terminal pty issue: Reproduced in a standalone PowerShell 7.6.4 window (not VS Code) — still times out identically.
• Corrupted/incomplete install: Full uninstall (Remove-Item on .local\bin\claude.exe and .local\share\claude) and reinstall via irm https://claude.ai/install.ps1 | iex, tested in a fresh terminal session and fresh Claude Code session (no -c resume) — still times out identically.
• Multiple GitHub scopes / other MCP config conflicts: Unrelated "github" server conflicting-scopes warning present throughout, but "github" (HTTP transport) connects fine in every test; only the stdio server is affected.

Expected Behavior

The "sqlcl" stdio MCP server should connect within a few seconds, consistent with its behavior under MCP Inspector.

Additional Notes

• Claude Code's debug log noted Using bash path: "C:\Program Files\Git\bin\bash.exe" at startup, suggesting some Windows process-spawning may route through Git Bash. This may be relevant to how stdio MCP child processes are spawned on Windows, though I have not been able to verify this directly.
• Happy to provide the full --debug log or test additional configurations if useful.

What Should Happen?

sqlcl should connect within seconds (as it does via MCP Inspector), not time out at 30000ms every time.

Error Messages/Logs

Steps to Reproduce

1. Create a .mcp.json file in a project root with a stdio-type MCP server pointing to a local native executable:

{
"mcpServers": {
"sqlcl": {
"type": "stdio",
"command": "C:\\sqlcl-latest\\sqlcl\\bin\\sql.exe",
"args": ["-mcp"]
}
}
}

(This example uses Oracle SQLcl's built-in MCP server mode, but the issue is not specific to SQLcl — see "What I've already ruled out" below.)

1. Start Claude Code in that project directory:

claude

1. Run:

/mcp

1. Observe that the "sqlcl" server shows "✘ failed" after approximately 30 seconds.

1. For comparison, verify the same command connects successfully outside Claude Code:

npx @modelcontextprotocol/inspector "C:\sqlcl-latest\sqlcl\bin\sql.exe" -mcp
→ Connects successfully in ~11 seconds via MCP Inspector's web UI, confirming the server itself and its MCP protocol implementation are working correctly.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.220

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

Note: I was only able to test this with the SQLcl binary I have available. The issue may reproduce with any native .exe configured as a stdio MCP server on Windows — I have not yet had a chance to test with a more universally available binary. Happy to test additional repro cases if the maintainers have a suggestion for one.