Vantage
Local-first analytics for Claude Code: token, cost, and agent-delegation insight from your own session logs.
thousands of sessions · zero data leaves the machine
- role
- Architect and sole engineer
- stack
- Next.js · TypeScript · Tailwind v4 · virtualized lists
- status
- oss
// 01 - PROBLEM
Heavy Claude Code users accumulate thousands of .jsonl session transcripts and no way to answer: where did the tokens go, what did each session cost, which projects and tools dominate? Existing tools ship your usage data to third parties.
// 02 - APPROACH
- Reads the session logs Claude Code already writes to ~/.claude/projects - no instrumentation, no telemetry.
- Insights dashboard: token totals, year heatmap, model mix, per-project and per-branch rollups, tool frequency.
- Session explorer: server-side search, sort, filter over a virtualized list; per-session agent timeline and delegation tree.
- Cost view: per-model estimates, 5-hour billing-window tracker, burn-rate meter.
// 03 - ARCHITECTURE
~/.claude logsparse oncemtime cacherollupsdashboard
- Parse once, cache by mtime
- Each file parses a single time into a compact per-session contribution; full message events are dropped after rollup. Memory stays bounded across thousands of sessions.
- Local-first, always
- Everything computes on the machine that owns the data. Privacy mode anonymizes names for screenshots and demos.
// 04 - PRODUCTION-GRADE
- Stays fast at thousands-of-sessions scale (windowed virtualization, server-side filtering)
- Editable per-model pricing so estimates track reality
- Keyboard-accessible controls, loading skeletons, empty states
- Nothing leaves the machine - verifiable by reading the code
// 05 - ARTIFACTS