MCP server processes leak on host after subagent/session termination

Bug description

MCP server processes spawned by Claude Code are not terminated when a subagent or session ends. Each process consumes ~44 MB of RAM and accumulates over time, leading to significant memory waste.

Reproduction steps

1. Configure an MCP server in Claude Code settings (e.g., mcp-postgres-full-access)
2. Use the Agent tool to spawn multiple subagents that use MCP tools
3. Wait for subagents to complete
4. Check running processes:

``bash
ps aux | grep "mcp-postgres" | grep -v grep | wc -l
``

Observed behavior

After a session with ~30 agent runs and subagents, 172 orphaned MCP processes remained on the host:

$ ps aux | grep "mcp-postgres" | grep -v grep | wc -l
172

$ ps aux | grep "mcp-postgres" | grep -v grep | awk '{sum += $6} END {print sum/1024 " MB total, " NR " processes"}'
1087.31 MB total, 172 processes

Each process is npm exec mcp-postgres-full-access postgresql://... consuming ~44 MB.

Expected behavior

MCP server processes should be terminated when:

• A subagent (Agent tool) completes or is cancelled
• A Claude Code session ends
• The parent claude CLI process exits

Workaround

Manual cleanup:

pkill -f "mcp-postgres-full-access"

Active agents will re-spawn their MCP servers automatically.

Environment

• Claude Code version: 2.1.87
• OS: macOS (Darwin 25.3.0)
• MCP server: mcp-postgres-full-access via npm exec
• Usage pattern: Agent tool spawning multiple concurrent subagents with MCP database access

Impact

• Memory: ~44 MB per orphaned process, 1+ GB after heavy sessions
• Connections: Each process holds an open PostgreSQL connection
• Accumulation: Processes never self-terminate; they persist until manual kill or system restart