Token Dashboard scans your local Claude Code JSONL transcripts hourly, deduplicates sessions by message ID, and computes exact USD cost using model-specific input, output, and cache token rates. Seven KPI tiles, a date range selector (7d / 30d / 90d / ALL), and six ECharts panels give you a full picture of your usage. Sessions, Projects, and Settings tabs round out the four-tab dashboard — all backed by Supabase, all derived from data that never leaves this server.
The Overview tab gives you 7 KPI tiles — Sessions, Turns, Input, Output, Cache Read, Cache Create, and Est. Cost — plus a date range selector (7d / 30d / 90d / ALL). Six ECharts panels cover: daily stacked token bars, daily cache reads, tokens by project (grouped bar), token share by model (donut), top tools by call count, and recent sessions. An expandable glossary explains every number. Every figure is computed from raw token counts using model-specific Anthropic rates.
The Sessions tab is a paginated table of every Claude Code session TKN has recorded — 25 per page, sortable by date, cost, message count, or tool call count. Each row shows the project slug, working directory, git branch, model, total tokens, cache hit percentage, USD cost, and session duration. Clicking a row expands an inline detail panel with the full token breakdown and all metadata fields. A search bar and model filter let you zero in on any specific session or project.
The Projects tab groups sessions by project slug — the directory name Claude Code derives from your working directory. Each row shows total USD cost, session count, aggregate token volume, average cost per session, and the last active timestamp. A horizontal bar chart makes the biggest spenders immediately visible. This is where you find which codebases are consuming the most compute budget.
The Prompts tab surfaces every distinct prompt extracted from your Claude Code sessions — the actual text you sent, de-duplicated and searchable. Each row includes the model used, session context, and token cost. A Save to PRL button on each row pushes the prompt directly into the PromptLibrary as a draft, so useful patterns discovered in sessions don’t get lost. A built-in duplicate detector groups identical prompts by hash and surfaces the estimated wasted spend from re-sending the same prompt — a fast way to spot patterns worth turning into a saved template.
Set a monthly token budget in the Settings panel and configure Telegram alert thresholds
at 50%, 75%, and 90% of spend. A dedicated run_tkn_budget_check.py script runs
hourly and fires a Telegram message the moment you cross a threshold — so you find out
before the bill arrives, not after. Register outbound budget
webhooks to POST a JSON payload to any endpoint when spend crosses your chosen percentage,
and wire up a weekly digest that recaps the last
7 days of cost by model straight to Telegram.
The Overview panel now includes a Cost by Model breakdown sourced from the MVS call log. Every Claude API call logged through the Model Vault Service is attributed by model, giving you a precise split of Opus / Sonnet / Haiku spend — not just per Claude Code session, but across all platform blueprints that route through MVS.
Every session in TKN originates from the JSONL transcript files Claude Code writes locally. A cron job runs hourly, deduplicates messages, computes exact USD costs, and upserts to Supabase atomically — no external calls to Anthropic required.
~/.claude/projects/**/*.jsonl hourly at :15. Each file maps to one Claude Code session.usage blocks containing input, output, and cache token counts.message.id. When the same ID appears in multiple writes, the highest output token count record wins.tkn_sessions via ON CONFLICT DO UPDATE. Dashboard KPIs and charts update from the live database.TKN uses Anthropic's published per-million-token rates for the Claude 4 and Claude 3 model families. Cache read tokens are priced at a deep discount relative to fresh input tokens — the savings accumulate significantly in long-running sessions with large context windows.
Seven KPI tiles, six charts, full session history, project breakdown, prompt extraction with duplicate detection, budget alerts and webhooks, a weekly Telegram digest, and model cost attribution — all from your local JSONL files.
Your Claude Code sessions are already writing JSONL transcripts. Token Dashboard turns those files into cost visibility — no API keys, no external calls, no configuration required.
Open Token DashboardRequires a MelTuc account. Create one free.