convez.ai
PluginProdutividadeIntermediário
iam25th1

Claude Context Pill

iam25th1/claude-context-pill

Contador visual de tokens para Claude.ai em tempo real

Ver no GitHub
BomPluginScore · de 0 a 100

Extensão Chrome que exibe um contador flutuante de uso de tokens na janela do Claude.ai. Usa tokenização real (js-tiktoken) para mensagens visíveis e estimativas conservadoras para overhead de sistema, arquivos e artifacts. Inclui modo de calibração opcional contra a API da Anthropic.

Instalação

$npx convez-ai iam25th1/claude-context-pill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 6 meses
Versão
v1.2.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Monitorar consumo de contexto durante conversas longas no Claude.ai
  • Evitar perda de contexto iniciando novo chat antes de atingir limite
  • Calibrar estimativas locais contra tokenização real da Anthropic

Veredito

Ferramenta prática e bem-engenheirada para quem trabalha com contextos longos no Claude.ai. A implementação com js-tiktoken real + calibração opcional mostra rigor técnico, mas depende de permissões de extensão e API key para funcionalidade completa.

README

A tradução para o português deste README ainda está na fila.

claude context pill

a tiny chrome extension that pins an estimated token-usage counter to the claude.ai composer.

license: mit version manifest v3 tokenizer


claude.ai does not expose a context-usage counter. this extension scrapes the visible conversation from the DOM, estimates tokens with a char-based heuristic, and renders a floating pill near the composer. it tells you, roughly, how much of the 1M token window you've burned.

  ●  ~248k / 1M   24.8%       <  60%   green
  ●  ~712k / 1M   71.2%       60-85%   amber
  ●  ~924k / 1M   92.4%       >  85%   red

why

long sessions silently drift toward the context ceiling. you only notice when claude starts forgetting things you said an hour ago. this pill gives you a visual signal so you can pre-emptively start a fresh chat, hand off context, or trim attachments.

install

from source

  1. git clone https://github.com/iam25th1/claude-context-pill.git
  2. open chrome://extensions
  3. enable Developer mode (top right)
  4. click Load unpacked and select the cloned folder
  5. reload any open claude.ai tab

the pill mounts bottom-right above the composer. the toolbar icon opens settings.

from chrome web store

coming soon.

how the estimate works

the pill ships a vendored copy of js-tiktoken with the cl100k_base BPE vocab. visible message text is tokenized with the real BPE algorithm, not a character heuristic, so the visible count is around 95% accurate against claude's actual tokenization.

the hidden portion (system prompt, skills, memory, tool definitions) is not observable, so the pill estimates it via feature detection:

featureoverhead contribution
baseline (vanilla chat)~18,000
project mode+22,000
each attached file+4,500
each web search performed+6,500
each artifact+3,000
each visible thinking block+1,500
user memory active (manual)+12,000

defaults are conservative starting points. the calibration mode (below) refines the visible-count accuracy against the real anthropic endpoint.

optional calibration mode

opt-in. uses your anthropic api key to validate the local tokenizer against the official /v1/messages/count_tokens endpoint.

  1. open the popup, switch to the calibration tab
  2. paste your anthropic api key
  3. toggle calibration enabled (grants permission for api.anthropic.com only)
  4. save

every Nth conversation change (default 10) the extension sends the visible text to anthropic's count endpoint. the ratio between the official count and the local estimate becomes a multiplicative correction factor, applied to all future estimates. ratio is clamped to [0.5, 2.0] so a bad response cannot poison the estimator.

the api key is stored only in chrome.storage.local. it is never logged and never sent anywhere besides api.anthropic.com. the optional host permission is dropped automatically if you toggle calibration off.

features

  • floating glass pill with click + drag + magnetic snap
  • snaps to any of 8 edges on release: composer top/bottom/left/right and screen top/bottom/left/right
  • position persists across reloads and follows the composer through sidebar toggles, window resizes, and route changes
  • 3-tier color status with subtle pulse animation
  • real BPE tokenization (no character heuristic)
  • smart per-feature overhead detection
  • click pill to toggle compact (~612k / 1M) and raw (612,348 / 1,000,000) views
  • settings popup with main + calibration tabs
  • opt-in api calibration with rolling correction factor
  • SPA route-change aware (switches conversations cleanly)
  • mutation observer with debounced re-render, low overhead
  • respects prefers-reduced-motion

runtime dependencies

zero. the vendored js-tiktoken bundle in vendor/tokenizer.bundle.js is the only third-party code that ships in the extension. no remote scripts, no remote fonts, no CDNs at runtime.

build (only needed if editing the tokenizer)

the extension loads vendor/tokenizer.bundle.js directly, so no build step is required to run or develop the rest of the extension. if you want to rebuild the tokenizer bundle from source:

npm ci --ignore-scripts
npm run bundle:tokenizer

this produces a fresh vendor/tokenizer.bundle.js from the pinned js-tiktoken version in package.json.

security

  • manifest v3, host scoped to https://claude.ai/* only
  • optional_host_permissions for https://api.anthropic.com/*, requested only when calibration is enabled and dropped when it is disabled
  • permissions limited to storage
  • no remote scripts, no eval, csp-locked extension pages (script-src 'self', connect-src https://api.anthropic.com)
  • message content is read from the DOM but never stored or transmitted (unless you opt into calibration, in which case it is sent only to the official anthropic endpoint)
  • api key, when supplied, is stored only in chrome.storage.local
  • calibration correction factor is mathematically clamped so a malformed remote response cannot poison the estimator

see SECURITY.md for the full policy and how to report issues.

privacy

see PRIVACY.md. short version: nothing leaves your browser.

contributing

PRs welcome for selector fixes, tokenizer improvements, ui polish. see CONTRIBUTING.md for the dev loop and test checklist.

roadmap

  • v1.3 shareable usage card export
  • v1.4 per-conversation profile-based overhead history (long-term learning)
  • v1.5 toolbar badge with live count

full history in CHANGELOG.md.

license

MIT, built by 25TH

Ver o original no GitHub

Bom para

Relacionados