Features
LLM spend FinOps: the pieces, each one real.
Tracking proxy
POST /v1/log records one LLM call (model, input/output tokens, feature tag) and returns its estimated cost. Stdlib only — no dependencies.
Per-feature cost attribution
Every call tagged by feature (JSON field or X-Feature header). The dashboard breaks spend down per feature with call counts and token totals.
Slack budget alerts
Unlimited alert rules on Starter: per-feature monthly caps, breach posts to Slack with spent-vs-cap and percentage.
Hard budget caps (Scale+)
When a feature is over its cap, /v1/log answers 429 and records nothing — your app treats the 429 as a stop signal and halts before its next provider call. Workspace-wide caps too.
Anomaly detection
Flags features whose spend today exceeds 3× their 7-day daily average (minimum $5/day to avoid noise).
Dashboard API
GET /dashboard: MTD spend, by-feature breakdown, budget breaches, anomalies — one JSON payload.
Retention windows
Plan targets: 7 days (Free), 90 days (Starter), 1 year (Scale). Automatic purge is on the roadmap — today, records stay until you delete them.
API-key auth
Per-workspace API keys, stored hashed — a key only ever sees its own workspace. With no keys configured, the endpoints refuse (fail closed). /health stays public.