kyde-score

ClaudeCommunityby Jürgen Eckel

Kyde Score watches how your AI agents behave and tells you the moment that behavior changes. Agents emit tool-call telemetry anyway — Claude Code, OpenCode and other OpenTelemetry-instrumented agents simply point their OTLP exporter at score.kyde.com, and Kyde Score turns that stream into a drift score per agent. Connecting is one click: sign in with your email and a one-time emailed link — no password, no API key to paste. Scoring keys you create are tied to your address, so your scores, alert webhooks and key management follow you across sessions, and critical actions (reset a baseline, delete a key) work only for you, never for an agent. How it works: the service learns each agent's baseline from its first 30 runs (tool mix, tempo, task contracts), then scores every run after that: 70–100 normal, 40–69 watch, 0–39 deviating. Detection is deterministic, not an LLM judge: CUSUM catches abrupt shifts, a divergence channel catches gradual drift, and task contracts flag any run that touches a tool or entity never seen in the baseline — even when the tool distribution looks perfectly normal, which is the signature of a hijacked agent. Same log, same score: every published score can be recomputed from the stored records. Tools: create_key (issue a scoring key, optional webhook), get_scores (current score and alarms per agent), set_webhook (be notified when a score changes band), claim_key (attach a human to a key for double-opt-in ownership), rebaseline (forget a learned baseline and relearn — human-only on claimed keys), revoke_key (delete the key and all data under it).