- Input
- OpenTelemetry over OTLP/HTTP, JSON or protobuf. Claude Code, Grok CLI, or any framework with GenAI tool spans.
- Setup
- Three environment variables, or one line to your agent: Set up score.kyde.com/skill.md
- Baseline
- Learned per agent over its first 30 runs and at least 3 days, then frozen and fingerprinted.
- Detectors
- CUSUM for sudden shifts, Jensen-Shannon divergence for slow drift, duplicate calls for loops (needs the parameter hash kyde.args.digest). Plus a task-contract check: a tool never seen for the task. No LLM.
- Score
- 0 to 100 per run. 70 and up normal, 40 to 69 watch, below 40 deviating. Counts the alarms of the last 20 runs, so an agent that settles recovers.
- Output
- Scores as JSON over the API and the MCP server. Signed webhook on every band change and every rebaseline.
- Data
- Call metadata only. Everything else is dropped before storage. Hosted in the EU. 90 days history.
- Effect on the agent
- None. Export is asynchronous. If Kyde is down, the agent runs as before.