[BUG] Remote MCP tool calls dropped at 60s client-side despite ▎ server responding in <10ms

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?

Remote MCP tool calls intermittently fail at the Claude Code
client side with "Remote tool call timed out" exactly 60 seconds
after dispatch. Server-side tool call logs prove the server
returned a successful response in single-digit milliseconds — so
the response is being dropped somewhere between the remote MCP
server and the Claude Code client, not on the tool's end.

Affects multiple tools on the same connector (not tool-specific):
economic_calculator_projections, run_read_query,
alumni_directory, degree_distribution, total_employment_records.
Observed rate ≈ 25% of calls. Retries within seconds usually
succeed.

Connector: claude.ai remote MCP connector to a FastMCP server
(Streamable HTTP transport) behind AWS ALB/ECS. Running under
Cowork dispatch child session.

What Should Happen?

When a remote MCP server returns a successful response in a few
milliseconds, the Claude Code client should surface that response
to the tool-use turn, not emit "Remote tool call timed out" 60s
later as if no response arrived.

Error Messages/Logs

---
  Error Messages/Logs (rendered as shell):

  # ~/Library/Logs/Claude/main.log — client side
  2026-04-17 07:04:16 [info] Making remote MCP tool call:
  2026-04-17 07:04:40 [info] [EventLogging] Flushing 16 events
  2026-04-17 07:05:16 [error] Remote tool call timed out

  # Server-side tool_call_logs row for the SAME call (FastMCP
  middleware logs
  # every tools/call after response status is determined; insert is
   fire-and-forget)
  created_at        | 2026-04-17T12:04:16.578687Z
  tool_name         | economic_calculator_projections
  parameters        | {"base_year": 2025, "growth_rate": 0.03,
                    |  "institution": "Ohio State University",
                    |  "projection_end": 2034}
  response_status   | success
  duration_ms       | 9
  error_message     | NULL

  # Four consecutive timeouts this morning all paired 1:1 with
  server-side successes:
  # local 06:39:42 -> server 11:39:42Z success 6ms
  # local 06:50:13 -> server 11:50:13Z success 8ms
  # local 06:51:13 -> server 11:51:14Z success 9ms
  # local 07:04:16 -> server 12:04:16Z success 9ms

  # Additional timeouts across unrelated tools in main.log history:
  2026-04-13 11:04:14 [error] Remote tool call timed out  #
  run_read_query
  2026-04-15 16:59:59 [error] Remote tool call timed out  #
  alumni_directory
  2026-04-16 10:53:20 [error] Remote tool call timed out  #
  degree_distribution
  2026-04-16 10:54:54 [error] Remote tool call timed out  #
  total_employment_records

  ---

Steps to Reproduce

1. Configure a remote MCP connector in claude.ai (OAuth,

Streamable HTTP).
Our server is FastMCP behind ALB/ECS at mcp.gradgraph.com.

1. In Claude Code (Cowork dispatch child), invoke any tool on

that connector
(e.g., economic_calculator_projections with
institution="Ohio State University").

1. Observe intermittently (~25%): client log emits

"Making remote MCP tool call: <name>"
followed exactly 60s later by
"Remote tool call timed out".

1. Query the remote server's request log for the same timestamp —

the request
was received and a success response was produced in <10ms
every time.

1. Retry the same call within seconds — usually succeeds in

5–30s.

Not tool-specific or parameter-specific. I initially suspected a
query-plan
issue on one tool; server logs disprove that (identical sub-10ms
durations for
"failing" and "passing" calls with the same parameters).

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.112

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_