First off, I call my OpenClaw assistant "Punk Ass," short for "Punk Assistant." It's been great so far having a punk ass for various projects. I hadn't been using my gaming machine much; now it's a work horse.
But, when you're running an AI assistant that handles Discord messages, cron jobs, and various projects; costs add up. But where exactly?
I directed a simple dashboard to find out.
How It Works #
OpenClaw stores session logs as JSONL files. Each assistant response includes token counts and costs. I had a Node.js script written that extracts this data, cross-references it with Discord thread metadata embedded in each conversation, and aggregates everything by thread, day, model, and session type.
The output is a single static JSON file that powers an Astro dashboard. No backend needed, just parse, deploy, and refresh when you want updated data.
What I Track #
I track a few things that help me make keep costs low:
- Cost per Discord thread
Which conversations are burning tokens? - Daily spend trends
Are costs stable or spiking? - Token breakdown
Input vs output vs cache hits - Session types
Discord vs cron jobs vs direct chat
What I Learned #
The data was immediately actionable. Two threads accounted for 60% of the total spend: a Shopify web store project ($73) and the #general channel ($70).
The store project made sense, I've been async art directing it alot. It's looking sweet. But #general? That's just a catch-all channel accumulating context from random conversations. Expensive by accident.
The Fix #
Now I know: long-running threads in Discord are context traps. Each message carries the full conversation history. The solutions:
- Start fresh threads for new topics instead of continuing in the same one
- Have OpenClaw delegate to sub-agents for self-contained tasks as they don't carry the parent context
- Treat
#generalas a lobby, not a workspace
The dashboard cost less than a dollar to make. But now, knowing where tokens go changes how I work with OpenClaw, I'll earn all that buck back quick; I can stop guessing and start optimizing.
