claude-code - 💡(How to fix) Fix [BUG] Cowork MCP client times out long-running tool calls despite server progress notifications [2 comments, 2 participants]

Symptom (user-visible error from host):

  MCP error -32001: Request timed out


Reproducible via stdio tee shim wrapping mcp-remote between Claude Code
host and remote MCP server. All internal identifiers redacted.


=== Outbound JSON-RPC from host (in.jsonl) ===

{"method":"initialize","params":{"protocolVersion":"2025-11-25",
 "capabilities":{"roots":{"listChanged":true},
 "extensions":{"io.modelcontextprotocol/ui":{"mimeTypes":["text/html;profile=mcp-app"]}}},
 "clientInfo":{"name":"local-agent-mode-<server-id>","version":"1.0.0"}},
 "jsonrpc":"2.0","id":0}

{"method":"notifications/initialized","jsonrpc":"2.0"}
{"method":"tools/list","jsonrpc":"2.0","id":1}
{"method":"notifications/roots/list_changed","jsonrpc":"2.0"}

{"method":"tools/call",
 "params":{"name":"<tool-name>","arguments":{"message":"<REDACTED>"}},
 "jsonrpc":"2.0","id":2}
 ^^^ Note: params._meta is ABSENT. No progressToken supplied.

{"jsonrpc":"2.0","method":"notifications/cancelled",
 "params":{"requestId":2,
           "reason":"McpError: MCP error -32001: Request timed out"}}


=== Inbound JSON-RPC to host (out.jsonl) ===

{"jsonrpc":"2.0","id":0,"result":{"protocolVersion":"2025-03-26", ...}}
{"jsonrpc":"2.0","id":1,"result":{"tools":[...]}}

{"jsonrpc":"2.0","id":2,"result":{"content":[{"type":"text",
 "text":"<REDACTED — valid full response payload>"}]}}
 ^^^ Note: the result for the cancelled request id arrived intact AFTER
     the host issued notifications/cancelled. Host discarded it silently.


=== mcp-remote bridge stderr (err.log) ===

[pid] Connected to remote server using StreamableHTTPClientTransport
[pid] Local STDIO server running
[pid] Proxy established successfully between local STDIO and remote
      StreamableHTTPClientTransport
[pid] [Local→Remote] initialize
[pid] [Remote→Local] 0
[pid] [Local→Remote] notifications/initialized
[pid] [Local→Remote] tools/list
[pid] [Remote→Local] 1
[pid] [Local→Remote] notifications/roots/list_changed
[pid] [Local→Remote] tools/call
[pid] [Local→Remote] notifications/cancelled
[pid] [Remote→Local] 2
[pid]
Shutting down...


=== Independent verification: server emits compliant progress ===

Direct curl against the same MCP endpoint with an explicit progressToken
streams notifications/progress every ~5s with the matching token,
interleaved with SSE ": keepalive" comments. Sample frames:

  data: {"jsonrpc":"2.0","method":"notifications/progress",
         "params":{"progressToken":"probe-1","progress":1,
                   "message":"Processing your request..."}}

  : keepalive

  data: {"jsonrpc":"2.0","method":"notifications/progress",
         "params":{"progressToken":"probe-1","progress":2,
                   "message":"Still working on it, please wait..."}}

  ...

Server completes well within an extended client timeout (e.g.
`curl --max-time 300`). Confirms the transport, ingress, LB, and server
are healthy — defect is exclusively in the host's invocation of
client.callTool (no onprogress / no resetTimeoutOnProgress).


=== Versions ===

- mcp-remote: 0.1.37
- Transport: StreamableHTTPClientTransport, http-first strategy
- MCP protocol negotiated: 2025-03-26 (server) vs 2025-11-25 (client advertised)
- Host: Claude Code (via Cowork local-agent-mode wrapper; same code path)

---

#!/usr/bin/env bash
   set -uo pipefail
   LOG_DIR="$HOME/mcp-debug"; mkdir -p "$LOG_DIR"
   TS=$(date +%Y%m%d-%H%M%S)-$$
   exec tee -a "$LOG_DIR/in-$TS.jsonl" \
     | npx -y mcp-remote@latest "$@" 2>>"$LOG_DIR/err-$TS.log" \
     | tee -a "$LOG_DIR/out-$TS.jsonl"