Summary

MCP stdio servers using npx are spawned per-session but never cleaned up when sessions end. Over time, orphaned processes accumulate and consume significant RAM.

Observed Behavior

On a 30GB RAM server, 41 orphaned context7-mcp processes were found consuming ~2.4GB of RAM total (~60-100MB each). Each subagent/session spawn creates a new npm exec @upstash/context7-mcp process chain that persists after the parent session ends.

Timeline of accumulation:

• 19:30 → 6 processes
• 19:39 → +3 (9 total)
• 20:35 → +3 (18 total)
• 23:32 → +9 (33 total, burst from parallel subagent spawns)
• 23:49 → +3 (42 total)

Root Cause

1. MCP stdio protocol requires each client to spawn its own server process
2. OpenClaw gateway spawns a new MCP server per session/subagent invocation
3. When the session/subagent ends, the MCP server process is not terminated (no SIGTERM/SIGKILL sent)
4. npx exacerbates this by spawning a parent npm process + child node process (double the overhead)

Workaround

Installing the MCP server globally (npm install -g @upstash/context7-mcp) and changing the config from:

"context7": {
  "command": "npx",
  "args": ["-y", "@upstash/context7-mcp", "--api-key", "..."] 
}

to:

"context7": {
  "command": "context7-mcp",
  "args": ["--api-key", "..."] 
}

This reduces per-process overhead (~40MB vs ~100MB) but does not fix the orphaning issue. Processes still accumulate; they just use less RAM each.

Expected Behavior

The gateway should track child MCP server PIDs and send SIGTERM when the parent session/subagent exits, regardless of exit reason (completion, error, timeout, cancellation).

Environment

• OpenClaw version: 2026.4.14
• Host OS: Linux 6.8.0-110-generic (x64)
• Host: Hetzner VPS, 32GB RAM
• Session type: Telegram direct chat + subagent spawns

Related Issues

• MCP stdio servers accumulate across turns and are not cleaned up on config reload (memory leak) #60656 (same root cause, earlier report)
• Bundle MCP runtime never released for isolated cron sessions — unbounded process leak #69145 (cron session leak variant)
• MCP Server Resource Leak - Orphaned processes cause memory exhaustion and hanging commands #65694 (MCP resource leak)