Tokenscope MCP server

tokenscope ⏣

See what your AI-coding session actually cost — and what's eating your context. A local, read-only CLI that parses your Claude Code session logs and shows where the money goes: model output vs. context being re-sent every turn (the hidden 60%+ of most bills).

$ npx @wartzar-bee/tokenscope

  tokenscope ⏣  latest session
  ──────────────────────────────────────────────────────
  Total cost   $868.84   over 967 model turns

  Where the money went
  output (model writing)     ████░░░░░░░░░░░░░░░░░░░░  16%  $137.24
  cache read (re-sent ctx)   ████████████████░░░░░░░░  66%  $577.59
  cache write (new ctx)      ████░░░░░░░░░░░░░░░░░░░░  18%  $153.67

  Context size per turn  (peak 822k · avg 404k · now 822k tokens)
  ▁▁▁▁▁▂▂▂▂▂▂▂▃▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▆▆▆▆▇▇▇▇▇▇█

  Insights
  • Re-sent (cached) context cost $577.59 (66% of spend) — context re-read every turn.
  • Peak context ~822k tokens — /compact or a fresh session would cut per-turn cost.
  • Only 16% of spend is the model's actual output.

(A real session, default Opus pricing. Your numbers will differ — prices are overridable.)

Why

Agentic coding (Claude Code, etc.) produces surprise bills, and the cause is mundane: as a session grows, the whole context is re-sent every turn, so cost balloons even when the model writes little. Existing dashboards show totals; tokenscope shows the attribution — output vs. cache-read vs. cache-write vs. fresh input, the per-turn context-growth curve, cost by model, subagent spend, and which tools fill your context — with concrete "trim this" insights.

Try it in 10 seconds (no Claude Code logs needed)

npx @wartzar-bee/tokenscope --demo

Runs on a bundled sample session so you see the full report before pointing it at your own logs — no setup, nothing to configure. (The sample is synthetic, for demonstration.)

Install / run

No install — runs via npx:

npx @wartzar-bee/tokenscope               # your most recent Claude Code session
npx @wartzar-bee/tokenscope --demo        # a bundled sample session — no logs needed
npx @wartzar-bee/tokenscope --all         # aggregate every session
npx @wartzar-bee/tokenscope <file|dir>    # a specific session .jsonl
npx @wartzar-bee/tokenscope --version     # print the installed version and exit
npx @wartzar-bee/tokenscope --json        # machine-readable
npx @wartzar-bee/tokenscope --share       # privacy-safe shareable summary (markdown + SVG card)
npx @wartzar-bee/tokenscope --share-svg   # just the SVG "cost report card"
npx @wartzar-bee/tokenscope scan          # static token footprint of a source dir (the engine behind ci-guardrail)
npx @wartzar-bee/tokenscope scan --diff ../base   # cost delta of the current dir vs a base dir — catch a regression before you push
npx @wartzar-bee/tokenscope scan --max-total 50000            # exit 1 if the footprint exceeds a budget — a local cost gate
npx @wartzar-bee/tokenscope scan --diff ../base --max-delta 2000   # exit 1 if the diff adds more than N tokens

Reads ~/.claude/projects//.jsonl. Read-only, local, no network, no telemetry — open the source; nothing leaves your machine.

Local cost gate (pre-push / pre-commit)

--max-total N / --max-delta N make scan exit 1 when the token footprint (or a diff's delta) blows a budget — the same check ci-guardrail runs in CI, but locally, before you push. Wire the absolute budget into a git hook so a runaway prompt/config never leaves your machine:

Shortened. The full README is on GitHub.