Vinitokx

ViniTokx

Token analysis and optimization plugin for Claude Code and other AI coding assistants.

ViniTokx derives from the Sanskrit root "Viniyoga" — purposeful utilization.

Packages

Package	Description
@vinitokx/observer	Zero-token offline analyzer — reads Claude Code JSONL transcripts and reports per-model, per-tool token consumption without invoking any LLM. Ships the vinitokx CLI.
vtkxoptm	Rust binary that routes four narrow tasks (summarize, docstring, classify, boilerplate) to a local Ollama model so the cloud LLM never reads heavy input. Also provides vtkxoptm grep — semantic code search via local embeddings.
@vinitokx/claude	Claude Code plugin — slash commands (/token-analyze, /token-stats), the token-optimize skill, and the local-llm-runner subagent that auto-routes tasks to vtkxoptm.
@vinitokx/core	Token estimation engine used by the Claude plugin — heuristic char-based counting, section breakdown, and compression suggestions.
@vinitokx/cursor	Cursor IDE plugin (placeholder)
@vinitokx/codex	OpenAI Codex plugin (placeholder)

Getting Started

Step 1 — Install the local-LLM router

curl -fsSL https://raw.githubusercontent.com/gokul-kulkarni/vinitokx/main/packages/router/install.sh | bash

Step 2 — Install the Claude Code plugin

claude plugin add @vinitokx/claude

Step 3 — Init each repo you work in

cd your-project
vtkxoptm init

vtkxoptm init runs in one shot:

1. Checks Ollama is installed
2. Detects available RAM via llmfit and picks the best-fitting model, then pulls it
3. Warms the model so first inference is fast
4. Indexes the repo for semantic code search (vtkxoptm grep)

The plugin hooks (SessionStart / Stop) are wired automatically when you install the plugin — Ollama stays warm at the start of every session and your savings ledger prints when the session ends.

To change the model later: vtkxoptm setup. To re-index after big changes: vtkxoptm index.

Claude Code Skills & Commands

Name	Type	Trigger	Description
token-analyze	slash command	/token-analyze	Full per-model, per-tool breakdown from transcripts
token-stats	slash command	/token-stats	Single-screen ASCII dashboard
token-optimize	skill	"free up context" / "compress content"	LLM-driven compression with confirmation
local-llm-runner	subagent	auto-routed by Claude	Delegates summarize/docstring/classify/boilerplate to vtkxoptm

Development

bun install
turbo build         # Compile TS packages + build vtkxoptm Rust binary
turbo check-types   # Type-check all workspaces
turbo lint          # Lint all workspaces

License

MIT