Throttle

Pace control for AI limits.

Chrome extension that turns the opaque limits of Claude.ai and Lovable into measurable operational rhythm. Real-time consumption speedometer, depletion forecasting, historical sparkline, redline alerts. All data stays on your machine.

Supports Claude.ai and Lovable.dev. Multi-provider architecture ready for OpenAI and Gemini.

Purpose

Most usage tracker tools monitor passively. Throttle doesn't just measure — it controls the pace. The core metric is the PACE index:

PACE = current rate / ideal rate to reach zero at reset × 100

PACE 100 = perfect consumption, zeroes out at reset
PACE <50 = throttle is loose, can accelerate
PACE >130 = redline, will blow up before reset

This metric is stable across plans (Free, Pro, Max) because it's calculated on percentage of quota, not absolute tokens.

Architecture

Data capture: passive interception first

Instead of aggressively polling the /usage endpoint (anti-bot and rate limit risk), Throttle monkey-patches window.fetch in the MAIN world of claude.ai. Whenever Claude's own frontend calls the endpoint, the extension clones the response and processes it.

Flow:

┌─────────────────────────────────────────────────────────────┐
│ CLAUDE.AI (MAIN world)                                      │
│  ├── intercept.js → monkey-patches window.fetch             │
│  └── posts window.postMessage to ISOLATED world             │
├─────────────────────────────────────────────────────────────┤
│ EXTENSION (ISOLATED world)                                  │
│  ├── bridge.js     → forwarding layer                       │
│  ├── content.js    → renders Shadow DOM bar                 │
│  └── background.js → dedupe, analysis, alerts, storage      │
└─────────────────────────────────────────────────────────────┘

Active poll fallback (every 2min, configurable) runs only when passive interception goes silent — minimizes anomalous footprint.

orgId discovery via chrome.webRequest detects any call to /api/organizations/{uuid}/..., learning the org without requiring you to visit /settings/usage.

Shadow DOM isolated UI

The top bar is mounted inside a Shadow DOM. Changes to Claude.ai's CSS (redesigns, new themes) do not affect the extension. Zero style leakage risk.

Plan-agnostic storage

The snapshot schema renders only the buckets present in the payload:

{
  t: 1745350800000,
  u5h: 84.0,              // may be null
  u7d: 74.0,              // may be null
  reset5h: "2026-04-22T17:00:01Z",
  reset7d: "2026-04-23T20:00:00Z",
  extra_used: 15202,      // monthly in BRL
  extra_limit: 27500,
  extra_util: 55.28,
  extra_currency: "BRL"
}

If Anthropic changes the structure (adds an opus-specific field, renames five_hour to dynamic_window, etc), the tolerant parser ignores unknown fields and renders whatever exists.

Features

Top bar (24px, always visible)

THROTTLE (brand amber neon)
5H: % consumed, depletion ETA (or time to reset)
7D: % consumed, time to weekly reset
PACE: numeric pace index with color coding (green/yellow/red/blue)

Clicking the bar forces an immediate refresh.

Full popup

SVG speedometer (0–200, 100 = ideal) with animated needle and rev-limiter animation
Contextual legend that changes with status ("🟢 Healthy pace", "🔴 REDLINE · zeroes in 42min")
Claude window status (5h / 7d / extra BRL)
Lovable stats: daily and monthly usage side-by-side with historical sparkline
Separate ETAs for 15min and 60min
Full CSV export
Settings for poll interval, thresholds, and toggles

Chrome notification alerts

5h window crosses threshold (default 80%)
7d window crosses threshold (default 85%)
PACE enters redline (>130)
Stale telemetry (4h+ without data = possible endpoint change)

All alerts have cooldown to prevent spam.

Multi-account

Automatically detects organization switches via webRequest. Each org maintains separate snapshots. Dropdown in popup switches between them.

Installation

Clone this repo (or download it as a ZIP and extract it)
Open chrome://extensions
Enable "Developer mode"
"Load unpacked" → select the cloned folder
Open claude.ai and send a message — the bar appears at the top
Open lovable.dev to see Lovable credits in the popup
Popup available via the extension icon

File structure

throttle/
├── manifest.json              Manifest V3 with MAIN+ISOLATED scripts
├── intercept.js               MAIN world — monkey-patches fetch (Claude)
├── interceptor-lovable.js     MAIN world — intercepts Lovable fetch
├── bridge.js                  ISOLATED world — forwards postMessage
├── background.js              Service worker — analysis, alerts, storage
├── content.js                 ISOLATED — Shadow DOM bar (Claude)
├── content-lovable.js         ISOLATED — injection on Lovable
├── overlay.css                Placeholder (real styles live in Shadow DOM)
├── popup.html
├── popup.css
├── popup.js                   SVG speedo, sparkline, CSV export, settings
├── lib/
│   ├── storage.js             Multi-account, 8-day retention, dedupe
│   ├── predictor.js           Rates, ETA, PACE, status semaphore
│   └── providers/
│       ├── claude.js          Claude.ai parser/normalizer
│       └── lovable.js         Lovable.dev parser/normalizer
└── icons/
    ├── icon16.png
    ├── icon48.png
    └── icon128.png

Design decisions

Why PACE and not "tokens/min"

Absolute tokens change with the plan, the model (Sonnet vs Opus), and server load. Quota percentage is the only stable metric. PACE as a ratio (rate/ideal × 100) also decouples the interpretation from remaining time.

Why passive interception and not pure polling

Continuous polling of a private endpoint can trigger Anthropic's WAF. Claude's own application calls /usage multiple times per session — piggybacking on those calls eliminates anomalous footprint. Active polling only runs as fallback when the user is inactive.

Why Shadow DOM

Claude.ai has frequent redesigns. A <div style="position:fixed; top:0"> in 2026 may break in 2027. Shadow DOM ensures the host's CSS rules don't affect the extension and vice versa.

Why 8 days retention

The API's weekly window is 7 days. 8 days gives the sparkline room to cover the full cycle plus the reset transition.

Known limitations

Issue	Mitigation
Early-window (just after reset, sparse data)	Popup shows "Waiting for pace data" until 2+ snapshots exist
Granularity depends on traffic	Intercept captures every time Claude calls `/usage`; fallback poll guarantees minimum 1 snapshot/2min
Model used (Sonnet vs Opus) not differentiated	PACE measures aggregate effect; correlation with conversation type is left to the user
Private endpoint may change	Tolerant parser + "telemetry stale" alert after 4h without data
CSP block in MAIN world	Fallback: if monkey-patch fails, active polling continues working

Multi-LLM support

Provider	Status	Metrics
Claude.ai	✅ Active	5h window, 7d window, monthly extra
Lovable.dev	✅ Active	Daily and monthly credits, sparkline
OpenAI	Planned	`platform.openai.com/settings/organization/usage`
Gemini	Planned	`aistudio.google.com` quota pages

Each provider is a lib/providers/{name}.js module with discover() and normalize(payload). Adding a new provider requires no changes to the core.

Privacy

No third-party servers. The extension only talks to claude.ai and lovable.dev / api.lovable.dev (the services you already use), to read your own usage numbers.
No telemetry, analytics or tracking. Nothing is sent to the author or any backend.
No chat content reading. Only numeric usage endpoints are read.
Local storage only, in chrome.storage.local.

License

Unofficial project, no affiliation with Anthropic or Lovable.