Tokalator 🧮

Count your tokens like beads on an abacus.

A VS Code extension that shows you where your AI context budget is going — and helps you optimize it.

The Problem

AI coding assistants (Copilot, Claude, etc.) have finite context windows. A common complaint is "context rot," where performance degrades as conversations exceed a certain number of tokens. Opus 4.6 performs markedly better than its predecessors: on the 8-needle 1M variant of MRCR v2, Opus 4.6 scores 76%, whereas Sonnet 4.5 scores just 18.5%. This is a qualitative shift in how much context a model can actually use while maintaining peak performance.

Still, when you have 30 tabs open, your assistant's attention gets diluted across irrelevant files. You can't see what's happening. Until now.

Features

Token Budget Dashboard

A sidebar panel that shows your context usage at a glance:

Budget level — LOW (good), MEDIUM (warning), or HIGH (critical)
Next turn preview — estimated token cost before you send your next message
File list — ranked by relevance to your current task
One-click cleanup — close low-relevance tabs instantly
Pinned files — mark files as always-relevant (persists across sessions)
Session tracking — see last session stats on activation
High Contrast support — full contrastBorder handling for Dark, Light, and High Contrast themes

Chat Commands (`@tokalator`)

Command	Description
`@tokalator /count`	Current token count and budget level
`@tokalator /breakdown`	See where tokens are going
`@tokalator /optimize`	Close low-relevance tabs
`@tokalator /pin <file>`	Pin a file as always-relevant
`@tokalator /unpin <file>`	Unpin a file
`@tokalator /instructions`	List instruction files and their token cost
`@tokalator /model [name]`	Show or switch the active model
`@tokalator /compaction`	Per-turn growth and compaction advice
`@tokalator /preview`	Preview token cost before sending
`@tokalator /reset`	Reset session (clear turn counter)
`@tokalator /exit`	End session and save summary

Smart Relevance Scoring

Files are scored based on:

Same language as active file
Import relationships (active file imports this tab)
Path similarity (nearby in directory tree)
Recently edited
Has diagnostics (errors you're debugging)

What We Fixed

No fake precision — shows LOW/MEDIUM/HIGH instead of meaningless "67.3%"
Memory cleanup — token cache clears when documents close
Turn counter logic — read-only commands don't inflate the counter
Persistent pins — pinned files survive VS Code restart

Usage

Sidebar: Click the abacus icon (🧮) in the Activity Bar
Chat: Type @tokalator in Copilot/Claude Chat
Status Bar: Quick budget indicator in bottom-right

Settings

Setting	Default	Description
`tokalator.relevanceThreshold`	`0.3`	Tabs below this are marked low-relevance
`tokalator.windowSize`	`1000000`	Context window size in tokens (1M for Opus 4.6)
`tokalator.contextRotWarningTurns`	`20`	Warn after this many chat turns
`tokalator.autoRefreshInterval`	`2000`	Dashboard refresh interval (ms)

Known Limitations

Real BPE tokenizers for Claude and OpenAI; Gemini uses heuristic (~1 token per 4 characters)
Import parsing uses regex, which misses some edge cases (multi-line imports, dynamic imports)
Relevance weights are not scientifically tuned
Conversation cost estimated at ~800 tokens/turn (varies by message length)

Requirements

VS Code 1.99+
GitHub Copilot or similar AI extension (for chat features)

License

MIT

Tokalator

Vahid Faraji