Daily-total dashboards hide runaway agents. A cron that starts burning $12 per hour at 3am is invisible in a today-total until morning. Cost per hour is the metric that catches it. Here's how the shape works, what to measure, and how ClawMetry surfaces it for every runtime you run.
Daily totals average away the shape. A workload that ran quietly for 20 hours and then spiked to $200 per hour for the last 4 shows up as a $180-per-day account with no visible incident. Per-hour spend is the shape you need to see: it exposes the spike as a spike, and the spike is where the runaway loops, the mis-routed models, and the forgotten cron jobs live. Every serious observability platform in adjacent categories (Datadog for infra, Sentry for errors, Honeycomb for latency) surfaces the same shape at minute-or-hour resolution because averaging is where signal dies.
First: tokens per minute per runtime. A number you can eyeball. Second: cost per minute per runtime, priced with cache-awareness (the largest cost differential today is prompt-cache hit vs miss, not one model over another). Third: cost per turn, so you can spot the one turn that fanned out to five providers and cost twenty-five times the median. Get all three in one view, filterable by runtime, and every anomaly stops being anomalous.
Pick a cap that is 3-4x your normal per-hour spend under load. That gives normal spikes room and only fires on real anomalies. The alert should page whoever is on call for the agent, not the account owner (agents run in teams). ClawMetry sends the alert on Slack, phone, or webhook, and the approval queue pauses the offending session before its next call so the alert is a stop, not just a notification.
Per-node, per-runtime, per-session, per-hour spend, sliced any way. Cache-read vs uncached-input tokens per call, so a session with a low cache hit rate is flagged (prompt caching is the largest cost lever on Anthropic and OpenAI). Per-model routing shows when a session flipped models mid-run and what it cost. Cost per turn, sorted, catches the 90th-percentile turn that dominates the bill.
The Cost tab groups by runtime by default. OpenClaw and NemoClaw are free on every plan; the other 17 (Claude Code, Codex, Cursor, Aider, Goose, opencode, Qwen Code, Hermes, PicoClaw, NanoClaw, Pi, Deep Agents, n8n, Antigravity, GitHub Copilot, Grok, QM) share the same per-hour rollup, so you compare Claude Code against Codex against QM in one view. Every runtime page under clawmetry.com/runtimes/<slug> documents exactly which store ClawMetry reads and how cost is derived.
Start the free trial →