I burned $250 of tokens in a day with OpenClaw. That day turned into ClawPane.

I wired OpenClaw into Slack + Telegram for three “small” jobs:

• Morning digest
• Summarize long threads into action items
• Explain new errors and open GitHub issues

By lunch it felt perfect. By night I’d spent $250.

Not because OpenClaw is inherently expensive. Because I accidentally created a token snowball.

What actually happened

1) Context quietly grew until every call was heavy

Summaries were “last 50 messages,” then “include relevant prior decisions,” then “keep continuity.” Each run carried more baggage. I treated the agent like it had memory, but what it really had was an ever-growing prompt.

The kicker: the workflow still felt like it was doing the same thing. The cost curve didn’t.

2) Screenshots and tool payloads inflated requests

One workflow attached screenshots “for debugging.” Another pasted full tool outputs into the next step “just in case.” Every attachment and transcript became more tokens, and then those tokens became part of the next call too.

It wasn’t one big request. It was many medium requests slowly turning into huge ones.

3) Prompt caching didn’t behave like I assumed

I had a big system prompt (format rules, safety rails, tool guidance). I expected “write it once, reuse it forever.”

What I saw instead was behavior consistent with caching not sticking: repeated large prefixes being paid over and over. Even if you’re not hitting an obvious bug, relying on caching as your budget plan is fragile.

4) Idle gaps made the next run expensive again

My scheduled jobs created a pattern: idle for a while, then fire. If the next run has to re-establish a large prompt footprint, you keep paying the setup cost repeatedly.

That’s how “a few automations” became a token furnace.

The moment I decided to build ClawPane.co

What I needed wasn’t another usage chart. I needed control at the point where money actually leaks: model choice per request.

That’s what ClawPane is:

• It plugs into OpenClaw as a model provider and routes requests automatically.
• You stop hardcoding model names in every agent. You use router presets like auto, fast, economy, quality (or your own).
• You can tune tradeoffs per router (cost vs latency vs quality vs carbon).
• You get per-response visibility into what ran (selected model, cost, latency).
• You can define fallback behavior so one provider hiccup doesn’t cascade into retries and duplicated spend.

The thesis: keep OpenClaw workflows the same, but stop paying premium-model tax for routine steps and reduce blast radius when things get noisy.

The takeaway

OpenClaw bills don’t explode because one call is expensive.

They explode because:

• context grows,
• payloads inflate,
• caching assumptions fail,
• scheduled runs re-pay setup costs,
• and the system keeps running while you’re not looking.

If you want predictability, you need (1) hard limits and trimming, and (2) routing that chooses the cheapest model that still does the job.