agentscope renders one scan-first HTML page over your agent crew's scheduled jobs and activity — so you catch silent failures in ten seconds instead of discovering them in logs.
Python 3.8+ · no dependencies · data never leaves your machine
Why
LangSmith-class tools assume an ML engineer reading spans. But when your agent crew runs scheduled work around the clock, the human question is simpler — and almost nobody builds for it.
Config drift, expired credentials, accidentally disabled jobs — none of these raise their hand. They just quietly stop working.
The interface inverts: fewer buttons, more signal. agentscope has zero controls. It shows what was already done and whether anything's off.
KPIs up top, problems isolated in one panel, quiet when healthy. If the panel is empty, everything is fine — go make coffee.
Day one
"Two production jobs had been silently skipping every run for days. The scheduler noticed a model-config drift and skipped 'to prevent unintended spend' — with no error surfacing anywhere I looked. The dashboard's first render flagged both as FAIL with the exact fix hint. Fixed in two minutes."
— the build log, first real run
What you get
| Input | Your agent state: job definitions (JSON), execution history (SQLite), session directories. Config-driven — point it at anything. |
| Output | A single self-contained HTML dashboard. Dark or light theme. Opens on any device. |
| The panels | KPI row → Needs Attention (anomalies only) → scheduled jobs with ON/OFF → recent actions feed → sessions sparkline. |
| What it isn't | Not an orchestrator. Not distributed tracing. No multi-user permissions. No buttons at all — read-only by design. |
| Privacy | Reads local files, writes one HTML file. Nothing leaves your machine. |
# auto-detects Hermes Agent layout python3 generate.py -o dashboard.html # or point it at your own stack python3 generate.py --config myconfig.json -o dashboard.html