Observe. Evaluate. Debug.
Your OpenClaw agents.
The observability platform for OpenClaw, MyClaw, and KiloClaw. See every tool call, every token, every dollar — with auto-evals that tell you exactly what went wrong and how to fix it.
Free. No SDK. No signup. Drop your transcript or run an agent — traces appear instantly.
PostHog-level depth for AI agents
Every trace gives you seven layers of insight — from surface stats to prompt-level analysis.
Waterfall Timeline
Chrome DevTools-style bar chart. See exactly where time is spent — which tool call took 3 seconds, which one looped.
Cost Tracking
Cumulative cost sparkline with spike detection. Know which step blew your budget. Per-step token accounting.
Auto-Diagnosis
"Here's what went wrong" in plain English. Loop detection, refusal detection, hallucinated URLs, tool errors — 8 checks.
Fix Suggestions
Every failure includes the exact system prompt change to fix it. Copy-paste ready.
Prompt Analyzer
System prompt waste detection, duplicate sentence finder, constraint overload check, missing date warning.
Session Analytics
Grade trends, duration sparklines, model usage breakdown, cost per task type — across all your sessions.
Security Audit
SSRF detection, API key leak scan, PII exposure, prompt injection, dangerous commands, data exfiltration — 10 checks per trace.
Three ways to get traces
Drop a file
Drag your transcript.jsonl onto the page. Works with any OpenClaw, MyClaw, or KiloClaw session.
Run on AgentCub
Use our hosted OpenClaw agent. Every run auto-captures traces — waterfall, cost, evals appear instantly. No copy-paste.
Fastest path to insight →Share a trace
Save any trace and get a shareable link. Post it in Discord, paste it in a GitHub issue. Others see the full analysis.
Debug together →Built for the OpenClaw ecosystem
| AgentCub | LangSmith | Langfuse | |
|---|---|---|---|
| OpenClaw native | ✓ | ✗ | ~ |
| Zero setup / no SDK | ✓ | ✗ | ✗ |
| Auto-diagnosis + fixes | ✓ | ✗ | ✗ |
| Prompt analysis | ✓ | ~ | ✗ |
| Free tier | Unlimited | 5K traces | 50K events |
| Copy diagnosis for Discord | ✓ | ✗ | ✗ |
8 auto-evals on every trace
Rule-based. No LLM needed. Instant. A-F grading with fix suggestions.
Loop Detection
Same tool call 3+ times
Cost Alert
Over $0.10 or $0.50 threshold
Tool Compliance
Used web_search before answering?
URL Verification
Output URLs from tool results?
Efficiency
>15 tool calls or >3 min
Refusal Detection
"I cannot assist" patterns
Response Quality
Empty or trivial output
Tool Errors
Failed tool calls counted
The OpenClaw debug workflow
Someone in Discord says “my agent is looping.” Here's what happens:
They drop their transcript.jsonl at agentcub.live/observe
Instant diagnosis: "Agent called web_search 7 times with identical arguments"
Fix suggestion: "Add to system prompt: Never retry a tool call with the same arguments"
Click 'Copy for Discord' — paste the formatted diagnosis back
Share the trace link — everyone sees the full waterfall, cost, and timeline
Free: OpenClaw Production Config Cheatsheet
Every config key, every gotcha, one page. 15+ production configs I discovered the hard way.
Also included
13 interactive reveals across 3 tiers. Understand how the lobster works, layer by layer.
20+ articles from hosting OpenClaw in production. Every wall hit, every fix documented.
Browse and understand SKILL.md files. See how agents learn new capabilities.
See what your agent is actually doing.
Drop a transcript or run an agent. Traces in seconds.