Agent acp: MCP tools/call times out at ~60s with no way to configure it

Support Help
, ,
1

Summary

When using agent acp (Agent Client Protocol), MCP tool calls that take longer than ~60 seconds fail with:

MCP error -32001: Request timed out

There is no user-facing setting to increase this limit.

Environment

  • agent CLI: 2026.06.19-20-24-33-653a7fb
  • Cursor IDE (installed alongside): 3.8.22
  • Transport: MCP over HTTP (streamable)

Steps to reproduce

  1. Start agent acp.
  2. Connect an ACP client and create a session with any MCP server that exposes a long-running tool (any tool whose handler blocks for more than ~60 seconds before returning a final tools/call result).
  3. Invoke that tool through the agent.
  4. Wait past ~60 seconds.

Expected

  • The MCP client should wait long enough for legitimate long-running tools to finish, or
  • Provide a documented, user-configurable timeout for MCP tools/call (per server or globally).

Actual

  • The call fails at approximately 60 seconds with MCP error -32001: Request timed out.
  • Progress notifications from the MCP server do not appear to extend the wait window.
  • No configuration option exists in agent acp, ~/.cursor/mcp.json, or Cursor settings.

Notes

  • This affects the ACP / cursor-agent CLI path specifically (agent acp), not only the IDE.
  • The IDE MCP stack appears to use a much longer internal timeout for tools/call, so behavior is inconsistent between ACP and the editor.
  • Error text sometimes appears duplicated in tool results (e.g. ... Request timed out. ... Request timed out), which suggests layered error wrapping.

Request

Please either:

  1. Make MCP tool-call timeout configurable for agent acp / cursor-agent CLI, or
  2. Align ACP with the IDE on a reasonable default (e.g. 10–60 minutes for blocking tools), and/or
  3. Honor resetTimeoutOnProgress per the MCP SDK so servers can keep long calls alive with progress updates.
5

Hey, thanks for the detailed report. It’s super clear and really helps.

This is already a tracked bug, and there’s an earlier thread with the same symptom: ACP MCP client should reset JSON-RPC tool-call timeout on notifications/progress (or honour resetTimeoutOnProgress). I can confirm what you’re seeing. On the ACP/CLI path, tools/call uses the hardcoded SDK default of 60s, and notifications/progress don’t extend it. In the IDE, the timeout is noticeably longer. There’s no configurable setting for this right now. I can’t share an ETA for a fix yet, but I’ll post an update once I have one.

For now, here are a couple workarounds:

  • Make the long-running tool return a final result within about 60s. For example, return a job id quickly, then poll status in a separate call.
  • If it fits your scenario, run the long tool from the IDE. The tools/call timeout there is much longer, around 60 minutes.

Let me know if neither option works for your case.