Problem

On macOS, the default launchctl limit maxfiles soft limit is 256. Hermes gateway processes with multiple profiles + MCP servers routinely exceed this:

trade-paper gateway: 286 fds (198 IPv4, 50 REG, 20 PIPE)

When the limit is hit, any file operation fails with:

OSError: [Errno 24] Too many open files: '.sessions_*.tmp'

This crashes session saves, kanban dispatches, and any tool that opens files.

Root cause

• Each MCP server subprocess (playwright, sequential-thinking, memory, brave-search) maintains persistent pipes/sockets
• Multiple profiles each run their own gateway with independent MCP connections
• Hermes never closes idle MCP connections or monitors fd usage
• No startup warning when ulimit -n is near capacity

Workaround

Manually add SoftResourceLimits → NumberOfFiles: 4096 to each gateway launchd plist.

Suggested fixes

1. Startup detection: Check resource.getrlimit(resource.RLIMIT_NOFILE) at gateway start, warn if soft limit < 1024
2. MCP lazy init: Don't connect all MCP servers at startup — connect on first tool call
3. Idle MCP timeout: Close MCP connections idle for N minutes, reconnect on demand
4. Periodic fd check: Log warning in gateway loop if fd count exceeds 80% of soft limit

Environment

• macOS 26.5 (Apple Silicon)
• Hermes 3 profiles (default, research, trade-paper)
• 6 MCP servers configured
• Python 3.x, launchd-managed gateways