LOCAL · OFFLINE · CLAUDE CODE + CODEX

Know what your attention bought.

Cadence reads the Claude Code and Codex logs already on your Mac and turns them into one offline dashboard: hours of your attention, hours of agent work, what actually shipped — and the leverage between them.

×6.8leverage · today

77 %first-pass

59 %5h window util

See how it works

The Cadence Overview rendered from sample data: leverage ×6.8 (up 47.3 %) with the verdict '65 hours of attention conducted 440 hours of agent work at allegro tempo', the attention-vs-agent-work area chart, three insight cards, and the chain band — Attention 65 h, Leverage ×6.8, First-pass 77 %, Shipped 399 merged PRs. — The dashboard itself, on loopback. sample · jul 7 – 13

100 % offline by default loopback-only dashboard prompts never leave the box Python 3.11+ brew or pip

Up and reading in about 30 seconds

From logs to first readout in three commands.

brew install lubosxyz/tap/cadence
cadence scan
cadence serve

Prefer pip? pip install cadence-report · Python 3.11+ · Opens a local dashboard on your machine. No account required.

The dashboard

One dashboard, the whole loop.

Attention → leverage → first-pass → shipped. Every view is rendered from your own logs, on your own machine.

Process

Active, wasted and idle minutes against the 5-hour billing window.

Allocation

Where the hours went — by project, by tool, by session.

Score

Your day read like a metronome, largo to presto.

01 — THE CHAIN

One chain. Four numbers.

Attention is the scarce input. Cadence follows it through delegation and friction into shipped code.

Tempo is your 5-hour billing-window utilization, read like a metronome — largo under 20 %, adagio, andante, allegro, presto above 75 %. sample data · jul 7 – 13

ATTENTION

65 h

10 min per merged PR

The hours you were actually at the keyboard with an agent.
LEVERAGE

×6.8

440 agent-h from your 65 h

Agent-hours per attention-hour; the multiplier your delegation earns.
FIRST-PASS

77 %

678 clean · 201 one fix · 1 loop

How often work landed without a rework loop.
SHIPPED

399 PRs

API-list reference $1.40 each

Git-authoritative merged PRs, net LOC, commits.

02 — THE SCORE

Read your day like a conductor's score.

You on the top staff. Every agent session on a lane below. PRs, merges and rework loops marked where they happened. The strip underneath counts how many lanes were playing at once.

The Score for Thursday 9 July: a 00:00–24:00 timeline with your attention as ember blocks on the top staff, four agent lanes below carrying session bars with PR-opened and merged glyphs and one rework-loop mark, and an 'in play' strip showing three lanes at peak. — **3 lanes at peak** — the strip counts sessions playing at once ◇ PR opened · ◆ merged ‖ marks a rework loop Each lane's colour is the model that did most of the work

Open a sample score

03 — TWO LEDGERS

Two ledgers, honestly joined.

Session logs tell you how you worked. Git tells you what shipped. Cadence keeps them in separate ledgers and bridges them on one page — every panel is chipped with its source.

PROCESS · session logs

The Process tempo hero rendered from sample data: tempo Allegro at 59 % window utilization with 64h 36m active and 33h 27m wasted, the billing-window wave for Jul 7 – Jul 13, and the largo-to-presto tempo scale with the 'now' marker at 59 %.

active · wasted · idleminutes inside every session, read against the 5h billing window
5h billing-window utilizationtempo — largo · adagio · andante · allegro · presto
prompts / day · turns / prompthow you drive the agent, day by day
tokens & cache hitcontext spend and the reuse ratio behind it
project allocationwhere the hours went, per project and per window
agent & model mixClaude Code · Codex · per model, per day

SHIPPING · git

The Shipping ledger rendered from sample data: net lines +370.3k, 748 commits, 277 merged PRs, 15 repos touched, and the cumulative shipping ledger chart for Jul 7 – Jul 13.

net LOC · commits · merged PRsread from git, not guessed from chat volume
repos & languages · workday heatmapsince first commit, every clone on this machine

CORRELATIONS — hours vs output per project · tokens vs shipped, daily · process and git heatmaps side by side

The 'Workday shape — Process vs. Shipping' panel from the Correlations view, rendered from sample data: the same hour-by-weekday grid twice, session active minutes beside git commits, each with its own source chip. — Correlations, rendered — the only view that mixes both sources, and it says so on every panel. sample · jul 7 – 13

Two sources, two numbers, both kept: in the sample window the chain counts 399 merged PRs from PR metadata, while the git ledger counts the 277 merges it can see in the clones on this machine.

Act on the signal

Cadence tells you what to tune next.

The dashboard surfaces the patterns that are expensive to feel and hard to remember.

Rework loops

See where repeated correction consumed attention.

‖ rework loops, marked on the score and counted per range
Expensive tools

Find tools and models taking a disproportionate share of tokens.

“Model spend outweighs outcomes”
Focus drift

Spot fragmented time across too many projects or concurrent sessions.

“Too much time is fragmenting into idle mini-sessions”
Cache efficiency

Track the cache-hit ratio before context spend quietly compounds.

“Cache hit ratio is below healthy range”
Plan ROI

Compare a $200/mo Claude Max plan with API-list-cost equivalent usage and shipped output.

“Plan value cleared break-even”

4.75 projects

touched inside the same 5-hour billing window, on average — across 24 distinct projects in the range.

Spend less time reconstructing the week. Spend more time changing the next one.

04 — PLAN VALUE

Is the plan paying for itself?

Cadence imputes what your usage would cost at API list price and sets it against the plan you actually pay for.

Cadence uses your configured plan price — $200/mo Claude Max by default, or your own — and prices tokens at the providers' published list rates. Nothing is invented per token. sample data

CELEBRATE

Plan value cleared break-even: imputed usage was 11.94 × the configured plan cost.

WARNING

Model spend outweighs outcomes: claude-opus-4-8, 98.2 % of cost for 4.3 % of outcomes.

cache hit97.4 %

per merged PR at list$1.40

05 — EVERY DAY, EVERY MONDAY

Lives in your menu bar. Reports every Monday.

The macOS companion starts the local dashboard, keeps it fresh, and taps you on the shoulder once a week.

headline metric in the menu bar — today's leverage, first-pass, shipped or LOC touched
one-click rescan
scheduled scans — LOC pulse hourly, sessions daily, git weekly
first-run setup guide & launch at login
Monday nudge → Weekly Wrapped
private backup of a sanitized cache to a repo you own

Signed download ships with the public launch — the CLI works today.

Install Cadence

Download for macOS

Monthly Wrapped for June 2026, rendered from sample data: ‘You shipped more with less attention.’ — 481 merged changes from 283 hands-on hours; attention −20 %, merged work +180 %. — Wrapped: weekly, monthly, yearly — a shareable PNG, rendered offline. sample · monthly wrapped

06 — THE NETWORK BOUNDARY

Privacy is the architecture, not a setting.

Nothing on this page is a promise about a server. There is no server.

Read how it works

Zero network calls during scan and render.
The dashboard binds 127.0.0.1 with a per-session token.
Never reads your source files — only the agent logs and optional git metadata.
One SQLite cache under ~/.cache/cadence; delete it any time, Cadence rebuilds it.
Backup strips every free-text column and verifies the target repo is private before the first byte leaves.

No account.
No telemetry.
No remote fonts or charts.
No prompt, reply or tool text ever leaves the machine.

Three opt-in paths are the only ones that ever talk to the network: cadence publish (aggregates only) · cadence scan-github and scan-git --github (gh-backed PR metadata) · cadence backup (to a private repo you own).

07 — THE LEADERBOARD

A leaderboard you can't buy your way onto.

Opt in with cadence publish and only aggregates leave the box: ROI, net LOC per active hour, PRs merged per week, focus ratio, unique projects, active days — plus the two totals behind them and the cohort you chose (country, industry, language, plan, agent).

a slice is published only once ≥ 50 GitHub-verified builders are in it
podium ranks stay held until you verify
the verified badge is the server's decision, never the client's
pseudonymous handles · league tiers up to Vanguard

View leaderboard

published only once ≥ 50 GitHub-verified builders are in this slice

Sample leaderboard rows (ROI, net LOC per active hour, PRs per week, focus): the top three podium ranks are held until verified; ranks four to eight show verified builders.
Rank	Builder	ROI %	LOC / active hr	PRs / wk	Focus %
1	●●●● held · podium	—	—	—	—
2	●●●● held · podium	—	—	—	—
3	●●●● held · podium	—	—	—	—
4	orbit·2a6e ✓	642	123.9	4.3	82
5	nova·0f4c ✓	587	118.5	4.1	79
6	pine·8e31 ✓	521	101.3	3.6	77
7	sol·5d7b ✓	493	96.8	3.3	76
8	river·a12f ✓	461	90.2	3.1	74

League tierEmergingSteadyAscendantVanguardsample rows

08 — WHY NOT…

WakaTime counts keystrokes. GitHub counts commits. Cadence counts leverage.

What each tool measures: Cadence versus WakaTime versus GitHub
Metric	Cadence	WakaTime	GitHub
Attention vs agent hours	✓ yes	— no	— no
Leverage	✓ yes	— no	— no
First-pass & rework loops	✓ yes	— no	— no
5h billing-window utilization	✓ yes	— no	— no
Cache efficiency	✓ yes	— no	— no
Plan ROI	✓ yes	— no	— no
Active hours	✓ yes	✓ yes	✓ yes
Net LOC	✓ yes	✓ yes	✓ yes
Merged PRs	✓ yes	— no	✓ yes

09 — TEAM

Every machine and account, one dashboard.

Profiles and devices roll up locally. Teams get aggregates without raw text — the same boundary, at scale.

Early access opens with the public launch.

Request early access

10 — INSTALL

Your first score in under a minute. Offline.

brew install lubosxyz/tap/cadence
cadence scan
cadence serve

or pip install cadence-report · Python 3.11+ · macOS companion ships with the public launchmacOS app from Releases

macOS companion