Personal AI observability

Claude usage dashboard

Tools / observability · 2026 · Active — running locally

Knowing what I'm actually spending

I use Claude Code daily — sometimes for hours. Without observability, I had no idea how many tokens I was burning, which projects were costing the most, or whether my workflow was efficient. So I installed a third-party dashboard, hit a wall when it rendered blank, fixed the bug, then customized the whole thing to match my portfolio and added true-live auto-rescan so it updates as I work.

The dashboard is a single Python script with zero dependencies — just stdlib (sqlite3, http.server, pathlib) and Chart.js from a CDN. It reads Claude Code's local JSONL transcripts at ~/.claude/projects/, parses them into a SQLite database at ~/.claude/usage.db, then serves an HTML dashboard on localhost:8080 with token usage, cost estimates, model breakdowns, hourly distribution, and session history.

What it shows me right now: 1,216 turns across 30 sessions, 1.09M output tokens, 86.34M cache reads, ~$72 in API-equivalent cost over the last 30 days — Sonnet doing 79% of the work, Opus handling complex sessions, Haiku for fast tasks. Top project by token volume: Downloads/Claude_Coding (where most of my dev work happens).

What I contributed

Critical bug fix — dashboard.py:745 referenced an undeclared variable cutoff. The error got swallowed by an outer try/catch in loadData(), killing all renders silently — blank stats row and 4 empty chart cards even though the database had real data. Single-line fix: changed the filter to use the same start/end variables as line 662. Took 3 minutes of debugging, instant unblock.
Dark theme retrofit — rebuilt the entire color palette: --bg: #070C14, --card: #0F1822, --accent: #C9A084 (gold), --steel: #1C6FA3. 8-tone donut palette for the model breakdown. Red→steel gradient for the peak-hour bar.
Typography — swapped the default fonts for Inter (body), Montserrat (UI labels), and Cormorant Garamond italic for the brand wordmark.
Header rebrand — "Claude Code Usage Dashboard" → FORGE / Claude Usage. Title tag, header strip, everything aligned with the FORGE AI brand.
True-live auto-rescan — built a new endpoint POST /api/scan-incremental that runs the scanner without deleting the database. The 30s auto-refresh now hits incremental scan first, then loads data — so live Claude session turns appear within ~30 seconds of being written.
Backend system metadata — extended get_dashboard_data() to return a system block with DB path, size, mtime, total turns/sessions, oldest/newest record, unique projects, JSONL count, model count. All surfaced in the header meta strip.
Default range — changed from 30d to 7d (more useful daily-view default).

The full data flow is: Claude Code writes JSONL to ~/.claude/projects/ on every session (always, regardless of whether the dashboard is running). The scanner parses JSONL into a normalized SQLite schema. The dashboard reads from SQLite, not from JSONL — keeping it fast even with hundreds of sessions. Cowork sessions aren't captured because they're server-side and don't write local logs.

One nuance worth noting: the cost estimates use Anthropic's API pricing as of April 2026. Pro and Max subscriptions are flat-fee, so the dollar figure represents "what the API would cost" rather than actual spend. Useful for understanding usage patterns, not for billing. I treat it as a relative metric — am I burning more this week than last? Which project is the heaviest? That's what matters.

Active — running locally · auto-rescan every 30s

Real screenshots from the dashboard running on localhost:8080 — dark theme retrofit, true-live auto-rescan every 30 seconds, reading directly from ~/.claude/projects/*.jsonl.

FORGE / Claude Usage dashboard — full overview with stats, daily tokens, and hourly distribution — **Full dashboard view** — 1,216 turns · 30 sessions · 5 projects · 3 models · $72.06 est. cost (30d)

Daily token usage chart — last 7 days, stacked bars for input, output, cache read, and cache creation — **Daily token usage · last 7 days** — cache reads dominate output tokens: RAG context retrieval at scale

Average hourly distribution chart — turns and output tokens by hour of day, with peak hours highlighted — **Hourly distribution · peak hours detection** — 13:00 and 22:00 CT spikes: most productive coding windows

Cost by model table — Opus, Sonnet, and Haiku with turns, tokens, and estimated API cost — **Cost by model · 7-day breakdown** — Opus for complex tasks, Sonnet workhorse, Haiku for fast turns

Recent sessions table — session ID, project, last active timestamp, duration, model, turns, tokens, and cost — **Recent sessions · last 30 days** — per-session drill-down with model used and API-equivalent cost

Cost by project and cost by project and branch tables — Claude_Coding leads spend, with HEAD vs main branch breakdown — **Cost by project & branch · where the time goes** — Claude_Coding ($74.94, 22 sessions, 1.06M output tokens) is the heaviest lift; Portfolio-main and Personal_Dashboard fill the rest. The branch view splits HEAD from main so I can see which branch I'm actually burning tokens on.