Codex Desktop on Windows leaves Node/MCP helper processes alive until memory pressure makes PC unresponsive

Summary

Codex Desktop on Windows sometimes leaves many Node/MCP/helper processes running after Codex work, which appears to grow memory usage until the PC becomes sluggish or unresponsive. I am not sure whether the trigger is heavy subagent usage, MCP/Node REPL tool usage, or a broader helper lifecycle issue, but Task Manager shows many Node.js JavaScript Runtime and node_repl.exe processes after Codex activity.

This looks similar to the previously reported Windows helper stack issue in #26327, but I am filing a fresh report because the behavior is still visible on Codex Desktop 26.616.3767.0.

Environment

• App: Codex Desktop on Windows
• Codex package: OpenAI.Codex_26.616.3767.0_x64__2p2nqsd0c76g0
• OS: Microsoft Windows 11 Pro
• OS version/build: 10.0.26200, build 26200, 64-bit
• Approx. physical memory: 64 GB
• Usage pattern: long-running Codex Desktop work, including multiple threads/subagents and MCP/tool-enabled sessions

Observed behavior

On 2026-06-19 around 06:28 local time, Windows Task Manager filtered for node showed many Codex-related Node processes. The machine was under substantial memory pressure and became difficult to use.

From the screenshots:

• Task Manager showed overall memory around 67-69%.
• Many Node.js JavaScript Runtime entries were present.
• Several Node processes were using very large memory individually, including approximately:
• 1,448 MB
• 1,197 MB
• 1,054 MB
• 1,010 MB
• 869 MB
• The same filtered view also showed many node_repl.exe entries.
• CPU and disk activity for many of these processes were low, so the issue appears to be retained idle helper processes rather than active work.

A later live diagnostic sample after the system had recovered/settled still showed repeated Codex helper processes under Codex:

Category            Count PrivateMB WorkingMB OldestStart
--------            ----- --------- --------- ---------------------
codex_app_server        1     205.5     257.5 2026-06-19 06:44:42
mcp_server_cjs          3      68.3     116.1 2026-06-19 06:45:02
mcp_server_mjs          3     203.1     154.3 2026-06-19 06:45:02
node_repl               4     188.8     254.8 2026-06-19 06:45:02
other_codex_related    11     865.7     894.5

Representative live process command lines included repeated helpers like:

node_repl.exe
node.exe ./mcp/server.mjs --stdio
node.exe ./mcp/server.cjs --stdio

These were all children of the Codex process tree.

Expected behavior

Codex should keep helper process count and memory bounded during normal Desktop usage.

In particular:

• Completed or closed subagents should not leave their MCP/Node REPL/helper processes alive indefinitely.
• MCP stdio servers should be reused or torn down after the owning session/tool/subagent completes.
• Idle helper processes should not accumulate until Windows becomes memory-constrained.
• Users should not need to restart Codex or manually kill Node processes to recover the machine.

Actual behavior

After Codex Desktop usage, Task Manager can show many retained Node/MCP/helper processes. Some of the Node processes remain idle while holding hundreds of MB to more than 1 GB each, creating enough memory pressure that the PC becomes unresponsive.

Suspected trigger

I do not have a minimal isolated reproduction yet. The issue seems correlated with normal Codex Desktop work that uses:

• multiple threads
• subagents
• MCP tools
• Node REPL/tool-backed sessions

Subagent-heavy work may make it worse, but I cannot confirm that it is the only cause.

Related reports

• #26327: Windows accumulates Node REPL/MCP/stdio app-server helper stacks during long multi-thread sessions
• #25015: app-server leaks MCP process stacks for subagents, causing linear memory growth on Linux
• #20883: project-scoped MCP process pool instead of per-session MCP
• #20349: duplicate MCP/helper process trees and multi-GB RSS during UI/debugging workflows

Diagnostic commands used

These PowerShell commands were used to inspect the live process state:

Get-Process node,node_repl -ErrorAction SilentlyContinue |
  Select-Object Id,ProcessName,CPU,WorkingSet64,StartTime,Path |
  Sort-Object WorkingSet64 -Descending |
  Format-Table -AutoSize

Get-CimInstance Win32_Process |
  Where-Object { $_.Name -in @('node.exe','node_repl.exe') } |
  Select-Object ProcessId,ParentProcessId,Name,CreationDate,CommandLine |
  Sort-Object CreationDate |
  Format-List

$procInfo = Get-CimInstance Win32_Process |
  Where-Object { $_.Name -in @('node.exe','node_repl.exe') -or $_.CommandLine -match 'Codex|codex|mcp|node_repl' }

$perf = @{}
Get-Process | ForEach-Object { $perf[$_.Id] = $_ }

$rows = foreach ($w in $procInfo) {
  $p = $perf[[int]$w.ProcessId]
  $cmd = [string]$w.CommandLine
  $cat =
    if ($cmd -match 'node_repl.exe') { 'node_repl' }
    elseif ($cmd -match 'mcp/server\.cjs --stdio') { 'mcp_server_cjs' }
    elseif ($cmd -match 'mcp/server\.mjs --stdio') { 'mcp_server_mjs' }
    elseif ($cmd -match 'app-server') { 'codex_app_server' }
    elseif ($w.Name -eq 'node.exe') { 'node_other' }
    else { 'other_codex_related' }

  [pscustomobject]@{
    Category = $cat
    Id = $w.ProcessId
    PPid = $w.ParentProcessId
    PrivateMB = if ($p) { [math]::Round($p.PrivateMemorySize64 / 1MB, 1) } else { 0 }
    WorkingMB = if ($p) { [math]::Round($p.WorkingSet64 / 1MB, 1) } else { 0 }
    Start = if ($p) { $p.StartTime } else { $null }
  }
}

$rows |
  Group-Object Category |
  ForEach-Object {
    [pscustomobject]@{
      Category = $_.Name
      Count = $_.Count
      PrivateMB = [math]::Round(($_.Group | Measure-Object PrivateMB -Sum).Sum, 1)
      WorkingMB = [math]::Round(($_.Group | Measure-Object WorkingMB -Sum).Sum, 1)
      OldestStart = ($_.Group | Sort-Object Start | Select-Object -First 1).Start
    }
  } |
  Sort-Object Category |
  Format-Table -AutoSize