Token Savings

Refresh and open the Token Savings dashboard: a self-contained dark page that shows how many tokens and dollars two Claude Code context tools save —

• claude-mem: a memory system. Each observation stores the tokens it cost to originally produce that knowledge; summed = "work captured". Recalling it later is nearly free, so saved = work − recall.
• Graphify: a knowledge graph over a repo. Answering a structural question via the graph costs a fraction of reading the whole corpus (e.g. ~79.7× reduction).

The dashboard auto-detects both tools. If one isn't installed or has no data, that panel shows a friendly "not detected" state and the other still renders. No demo or fake data is ever shown.

Safety: read-only except for writing files under data/ (savings.js, savings.json, history JSON). It shells out to the sqlite3 and graphify CLIs already on the machine — no npm install, no network, no fetch. Node 18+ required.

Steps

1. Locate the dashboard directory (the cloned claude-token-savings-dashboard repo, which contains collect-savings.mjs and index.html). Resolve in order:
   • If $TOKEN_SAVINGS_DIR is set, use it.
   • Else use the path the user gives you, or a known clone location.
   • If you can't find it, ask the user where they cloned the repo. Confirm the directory contains collect-savings.mjs before proceeding.

2. Refresh the data. Run the collector from the dashboard directory so its relative paths and config.json resolve correctly:

   node "$TOKEN_SAVINGS_DIR/collect-savings.mjs"

   This rewrites data/savings.js (which sets window.SAVINGS) and data/savings.json. It is fail-safe — on any error it still writes a partial file and exits 0, so it never blocks.

3. Open the dashboard. It opens straight from disk (file://), no server:
   • macOS: open "$TOKEN_SAVINGS_DIR/index.html"
   • Linux: xdg-open "$TOKEN_SAVINGS_DIR/index.html"
   • (Windows: start "" "%TOKEN_SAVINGS_DIR%\index.html")
4. Summarize the headline numbers by reading data/savings.json and reporting back to the user, in one or two lines:
   • Combined: combined.savedDollars (saved $) and combined.savedTokens.
   • claude-mem: cloudMem.observations, cloudMem.workTokens (work captured), cloudMem.savedDollars, cloudMem.efficiencyPct — or its notDetected reason if the panel is empty.
   • Graphify: graphify.reductionFactor (e.g. 79.7x/query) and graphify.savedDollarsPerQuery — or its notDetected reason if empty.

Honesty notes (keep these in the summary if relevant)

• All figures are framed against a "load everything, every time" baseline you'd never actually run — directional, not an invoice.
• claude-mem efficiency is capped at 99.9% (never a fake 100%).
• Any Graphify monthly projection is illustrative — Graphify has no per-query log, so usage volume is assumed (graphifyAssumedQueriesPerDay).

Optional: auto-refresh on every session

To regenerate the data automatically each time Claude Code opens in the repo, add a SessionStart hook to .claude/settings.json (replace the path):

"hooks": {
  "SessionStart": [
    { "hooks": [ { "type": "command", "command": "node \"$CLAUDE_PROJECT_DIR/PATH/TO/collect-savings.mjs\" >/dev/null 2>&1 || true" } ] }
  ]
}