ClawMetry for NVIDIA NemoClaw - Know what's happening inside your NemoClaw sandboxes

Token tracking per session is exactly what I've been wanting. Running multiple sandboxes and having zero visibility into which ones are burning through credits is so frustrating.

The E2E encryption part is a nice touch too. Most monitoring tools want you to ship all your data to their cloud which is a nonstarter for anything sensitive. Open source MIT makes it easy to just try it without committing.

@vivek_chandI've been running a few AI agents myself , and seriously , most of the time I have no idea what's happening when one hangs or takes too long 😅.

The token tracking + history feature is a lifesaver ... I did not realize how unpredictable costs could get until I started monitoring them.

One idea I'd love to see: a quick health alert .. when an agent behaves differently than usual. Even a small

notification would save a lot of time checking manually.

@vivek_chand Many congratulations on yet another exciting launch, Vivek.

The one-command install with zero config is brilliant, reminds me of the early Docker days when simplicity actually mattered :)

Excited to see everything on the roadmap you’d shared with me.

@emma_watson21 Thank you Emma!

@nuseir_yassin1 Nuseir, really appreciate the support!

ClawMetry runs on both the host and inside each sandbox. The host side can observe policy changes as they happen; the sandbox side sees what the agent does within those boundaries. Policy drift detection on the roadmap will use exactly that to compare the active runtime policy against the static baseline and alert when they diverge.

@iamanantgupta Exactly that use case! The Flow tab shows you in real time where your chain is stuck and which tool call is hanging. No more guessing.

@zerotox Thanks Kumar! Great question. Each sandbox gets its own SSE event stream with color-coded multiplexing, so there's no cross-contamination. Sandboxes push events through E2E encrypted tunnels to ClawMetry Cloud, and decryption happens in your browser on demand, so the server never sees plaintext agent data. Lightweight JSON payloads keep everything snappy even with multiple sandboxes streaming simultaneously.

@roopreddy Roop, thank you! And that's one of the best edge cases to stress test with.

The brain tab caps at 500 events client-side, oldest events dropping off first. Two independent filter axes let you cut through burst noise: filter by agent source AND by event type simultaneously (AND logic, all purely client-side, no server round-trips). So if a sub-agent fires 40 exec calls, you isolate that agent's web_search events in two clicks. The server batches on a 0.5-second poll cycle, so a burst of 50 tool calls in 10 seconds arrives as roughly 5 batches rather than 50 individual updates. The honest limitation: no DOM virtualization yet, so very high burst rates can cause choppy re-renders. That's on the improvement list. The initial brain history load pre-clips to the 300 most recent events so the page stays fast after long agent runs.

@syed_shayanur_rahman Both! You get aggregate session totals on the Tokens tab, and the Brain tab breaks down individual tool calls with their costs. So you can see both the big picture and drill into exactly which call burned the most.

@maks_bilski Thank you Maks! Appreciate the support 🦞

@flowghost Spot on, Manchit. That "why did the quality agent reject?" problem is exactly what the Brain tab solves. You see every decision point across all agents in your pipeline with full context. Would love to hear how it works with your setup! And thanks, we just crossed the 100K installs mark! 🎉