The $30/month stack

After the April 4 ban, the stack was rebuilt from scratch, with the core idea being to separate low-cost models and high-performance models by role.

• GLM 4.7: used for repetitive tasks like heartbeats, memory checks, queuing, and housekeeping
• Pricing: $9/quarter unlimited plan
• GLM 5.1 or Sonnet: used for tasks that require real reasoning
  • Polishing writing to match my tone
  • Reviewing a thread before replying
  • Producing readable drafts before showing them to a person

OpenClaw's model routing handles task-based routing, and the user only defines the rules for which task goes where. Most day-to-day operations are handled by the $9 model.

In practice, the agent monitors the community, picks out threads worth engaging with, drafts comments in my voice, and puts them in an approval queue. It also tracks and logs posts that get good engagement. There's no auto-posting — everything must go through pre-publish approval.

The setup itself is maintained with Claude Code, while the running agent is handled by OpenClaw, dividing the roles this way. This brought total operating costs down to under $30/month.