Dashboard

Dashboard

An interactive Streamlit view over your extracted data: cost breakdowns, the session-depth meta-analysis, per-model and per-category charts, and quota trends. Charts are Apache ECharts (via streamlit-echarts) on a dark theme, and most charts cross-filter the whole board on click (and the time trend on a date brush).

It is an optional extra (it pulls in Streamlit, Apache ECharts, pandas and NumPy), so it is not installed by the core tool.

Want to see it first? The public live demo runs this exact dashboard on a synthetic dataset — no install, no real data.

Install & launch

# editable checkout (uv)
uv sync --extra dashboard
prompt-analytics dashboard            # reads ./output by default

# straight from PyPI, no install
uvx --with "prompt-analytics-for-claude-code[dashboard]" prompt-analytics dashboard

# or pip
pip install "prompt-analytics-for-claude-code[dashboard]"
prompt-analytics dashboard

(Use double quotes around the [dashboard] extra — single quotes are passed through literally by cmd.exe and break the install.)

The dashboard command runs streamlit run on the bundled app and serves on http://localhost:8501 by default; if the port is taken it picks the next free one and prints the URL. Any extra flags are forwarded straight to Streamlit, so prompt-analytics dashboard --server.port 8599 changes the port; you can also set defaults in a .streamlit/config.toml — see the Streamlit docs.

By default dashboard refreshes the data first — it runs extract + snapshot

• local (heuristic) categorize on your ~/.claude history, then opens the board, so a fresh launch never shows stale numbers. Pass --no-refresh to skip that and render the existing CSVs as-is. The refresh is skipped on the demo dataset (CCA_DEMO=1). If you start an extract while the dashboard is open, the cache keys on the CSV mtimes — a browser refresh (or the sidebar Refresh data button) shows the new data.

Choosing the data directory

Resolution order:

1. CCA_DEMO=1 — the bundled synthetic demo_data/ dataset, with a "Demo data" banner. This is what the public live demo runs.
2. CCA_DATA_DIR — an explicit directory of CSVs.
3. --output-dir DIR on the dashboard command (exported to the app as PROMPT_ANALYTICS_OUTPUT_DIR).
4. Default: ./output.

# Linux / macOS
CCA_DEMO=1 prompt-analytics dashboard            # synthetic demo data
CCA_DATA_DIR=/path/to/csvs prompt-analytics dashboard
prompt-analytics dashboard --output-dir ./out    # a specific extract

# Windows (PowerShell) — the `VAR=x cmd` prefix syntax does not exist there
$env:CCA_DEMO = "1"; prompt-analytics dashboard
$env:CCA_DATA_DIR = "C:\path\to\csvs"; prompt-analytics dashboard
prompt-analytics dashboard --output-dir .\out

Remove-Item Env:CCA_DEMO    # unset when done

Pages

All charts share one visual identity: color semantics live in dashboard/theme.py (stable colors per token type, per model family, per category) and the ECharts render layer in dashboard/echarts.py. Pages price everything on the primary provider (the first in pricing.yml, anthropic by default); the multi-provider grid stays available in the CLI.

Most categorical charts emit a cross-filter on click (project / model / category) and the time trend filters a date range by brush. Two aggregate clicks join them — a prompts-per-session bar (Sessions) and a prompt-length bucket (Composition / Input) — so any of them narrows the whole board. When a click filter is active a badge appears with two drill-through buttons, "Explore prompts →" and "Explore files →", opening the two detail pages on the current selection, plus a Reset. (The persistent sidebar filters — date / model / project / category — are separate; Reset clears only the click drill.)

The Compare tab (after Optimize) is the before/after view: pick a pivot date (or one of the auto-detected config-change dates) and a comparison window (1 week / 1 month / full history — the windowed modes keep N days immediately on each side for a fair like-for-like read), then read the five workload-normalized ratios as Before | After bands. Below them a Composition shift section makes the output/context move concrete — the code, test and context-rent shares plus the per-line efficiency (tokens and cost to write a line of code), with prose-vs-code and loading-vs-rent donuts before vs after — followed by two average-based charts (cost/prompt by token type, output language mix). On the full-history window it is built from the same report the impact command prints, so the bands and the CLI table never drift — and it is averages and ratios only, never sums.

Page	Shows
Home	KPI cards + cost-split donut, then cost over time stacked by model (click a model → filter) with a Day/Week/Month toggle.
Usage	Cost by token type per day/week (date brush + click → date filter), cache-read headline + non-cache volume bars, subagent cost, and the spend punchcard (weekday × hour).
Models	Two donuts (token volume / cost by model), per-prompt cost distribution by model (box plot), and cost by model over time (Day/Week/Month toggle). Every chart emits the model filter on click.
Sessions	Horizontal cost pareto by project (click → project filter), a cost treemap (click a tile → Prompt Explorer), prompts/session (click a bar → filter the board to that session size), and a box plot of cost/session by model.
Session depth	KPI deep-vs-opener cost multiplier, a box plot of one prompt's cost by depth band, and small-multiples of the input-side token mix plus output — the differentiating analysis.
Composition	"Where your cost goes, by content" — four sections, each a band of 4 KPIs + 2 charts: Input (cost by category + prompt-length histogram, both click-to-filter), Output (prose-vs-code donut + code language mix), Context (loading-vs-rent donut + a tokens Pareto), and Files — a force-layout file cost-graph (files sized by reconciled context cost, coloured by language; edit-prompt satellites; project picker + drill on a file). Mirrors by-output / by-context / by-file. (The old Prompts tab folded in here and into the Prompt Explorer; the category charts need categorize — see Prompt Categories.)
Prompt Explorer	The prompt detail page: drill session → prompt with a cumulative-cost timeline and the full prompt text; a selected prompt links its edited files straight to the File Explorer. Respects the global cross-filters; reached from the "Explore prompts →" button and treemap tiles.
File Explorer	The per-file footprint table (project / language / kind / scope filters + search) and a per-file drill: who edited it (with the task), which sessions kept it in context, and the load + rent it cost. Reached from "Explore files →", a prompt's edited-file links, and the Composition file graph.
Optimize	The prescriptive story: an avoidable-$ headline and three cards (long-pause cache rewrites, compact-earlier lever, compaction-is-cheap reassurance).
Compare	Before/after a pivot date with a comparison window (1 week / 1 month / full history): the five workload-normalized ratios as Before | After bands, a Composition shift section (output/context shares + per-line efficiency + prose-vs-code and loading-vs-rent donuts before vs after), and two average-based charts (cost/prompt by token type, output language mix). Reuses the impact numbers on full history; averages/ratios only.
Quotas	Quota windows first (utilization gauges, trend with reset windows, countdowns — need snapshot), then the cost story: plan break-even and Cost via GitHub Copilot (AI credits: subscription + overage beyond the allowance).
How it works	The trust argument displayed: dedup, ccusage reconciliation, pricing-at-read-time, categorization — and the cost inputs (per-token rate grid, Claude plans, Copilot credit allowances).

Configuration (config.yml)

The dashboard reads config.yml from the active data directory. Create the default with prompt-analytics config init. One section so far:

features:
  categorization: true   # gate the category charts (Composition / Input) + filter
  prompt_text: true      # whether extract stored full prompt text
  quota_snapshot: true   # gate the Quotas page

Missing keys fall back to these defaults. A features flag set to false hides the corresponding page; it does not delete any data.

See CLI Commands for the same data in the terminal, and Codebase Guide for how the dashboard layer is wired.