Overview

Live snapshot of the unified registry

LIVE 0 req/min 0 events
Window db — 0 ledger rows —
✦ v2 · IcyCache console × Loom dash, fused

THE TOKEN
VOLCANO

Cache hits, context compression and the tool tax — measured on one unified registry, then drawn by hand. Every tile is a real call read from SQLite: nothing here is estimated.

◭ true 3D — drag it ⛰ contour lines ✎ rough style · 100% pencil

Registry · v_token_ledger

—
Tokens paid upstream—
Tokens avoided—
Baseline (paid + avoided)—
Latency avoided by tool cache—
Threshold (≥ 35%)—

Calls per bucket

24h

Latency p50 / p95

milliseconds

Outcomes

hits vs misses

Per-model breakdown

ranked by traffic

Upstreams & clients

live
01

Activity — the request mountain

Day × hour (UTC), every tile is a measured call; the ink strokes are the isolines.

Traffic volume 3D

drag to rotate
drag ↔ scroll

Drag the mountain to spin it, hover a tile to read the hour.

Topo map

same surface

One curve = one traffic step.

Hourly traffic

hits vs misses per bucket

Latest requests

last 40 events
02

Token economy

Three domains, one unit: tokens actually paid versus tokens actually avoided.

Daily curve — paid vs avoided

14 days

Cross-hatched area = avoided tokens, dashed line = average, the record is circled in felt pen.

Economy gauge

against the 35% line

Where the savings come from

share of the baseline

Domain split

registry view

Economy per day

live
03

Cache — every stored prompt

Normalized prompt, hash and hit counter — straight out of the single SQLite file.

Most reused entries

hatched ranking

Cache entries

—
04

Context drawer — the memory

Sessions, sources, chunks, anchors and invariants kept by the context drawer.

Stored sources

before → anchor

Sources per session

live
05

Tool relay

Downstream MCP servers, their cache and the latency that never had to be paid.

Tool servers

gauge

Calls per tool

hatched bars

Tool cache

entries & bytes

Only tools on the cache_tools allow-list are cached — deny by default.

Per-tool breakdown

live
06

Audit live

The journal, breathing: every call and every relayed tool, as it lands.

LLM events

ledger journal

Tool events

relay journal
07

Guide — how to read this console

What the numbers mean, where a request travels, and which knobs to turn.

The path of a request

client → gateway → cache → upstream → registry → console

Reading the metrics

  • Hit — the answer came from the cache, so those upstream tokens were avoided, never billed.
  • Miss — the call went upstream; its tokens are real (upstream usage), never estimated.
  • Gain = avoided ÷ (paid + avoided). The operating threshold is 35%.
  • p95 — 95% of calls are faster than that.
  • A negative saving is possible: compression can cost more tokens than it saves. It is shown as measured.
  • Everything on this page is read from v_token_ledger, the unified registry.
  • All timestamps are UTC — one single clock, from the database to the charts.

The 13 MCP tools

Shortcuts & tips

  • 1…8 jump to a view, R refreshes, P pauses, E exports, C compacts.
  • The side panel switches view: every view is live, none of them is a static page.
  • The 3D mountain spins with the mouse; the map is read with the finger.
  • Ports: gateway 8150, MCP HTTP 8151, SSE 8152, console 8153.
  • Everything binds to 127.0.0.1 — nothing leaves the machine.
  • Compact in the side panel shrinks everything for a laptop screen.
  • Export downloads the exact JSON the page is drawing.

Live system

read from the admin API