👁 TokenLens AI

Intelligent token middleware for AI developers — see every token, control every request, pay only for what matters.

TokenLens AI is a VS Code extension that sits between you and any LLM. It automatically classifies your task, strips irrelevant context, routes to the right model, and shows you exactly what's happening with your tokens in real time.

The Problem

AI coding tools send your entire codebase to the LLM blindly.

A simple bug fix sends 50,000 tokens when 5,000 would do
You're paying 10x more than necessary
You have zero visibility into what's being sent

The Solution

TokenLens AI acts as a middleware layer that:

Classifies your task (bug fix, review, documentation, etc.)
Optimizes context — removes irrelevant files, compresses low-priority ones
Routes to the cheapest model that can handle the task
Shows you real-time token usage, cost, and savings

Your Prompt + Codebase (50,000 tokens)
              │
              ▼
        TokenLens AI
   ┌────────────────────┐
   │ 1. Classify task   │  → "bugfix"
   │ 2. Strip context   │  → remove 47 unrelated files
   │ 3. Compress files  │  → summarize 3 large files
   │ 4. Route model     │  → Claude Sonnet (not Opus)
   └────────────────────┘
              │
              ▼
    Optimized Request (6,200 tokens)  → 88% saved

Features

Token Dashboard — real-time view of tokens used, cost, and savings per request
Auto Task Classifier — detects bugfix, review, architecture, docs, refactor, test, and more
Context Optimizer — removes irrelevant files, compresses low-priority content
Smart Model Router — auto-routes to Haiku, Sonnet, Opus, GPT-4o, or local Ollama
Prompt Caching — leverages Anthropic/OpenAI cache APIs for stable context
PII Detection — warns before sending sensitive data to cloud LLMs
Stack Templates — pre-configured context profiles for popular frameworks
MCP Server Templates — ready-made integrations for GitHub, Jira, Adobe Commerce, and more

Quick Start

1. Install

Search TokenLens AI in VS Code Extensions or:

code --install-extension tokenlens-ai.tokenlens-ai

2. Configure API Keys

// VS Code Settings (settings.json)
{
  "tokenlens.anthropicApiKey": "sk-ant-...",
  "tokenlens.openaiApiKey": "sk-...",
  "tokenlens.defaultProvider": "anthropic"
}

Or set environment variables:

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

3. Select Your Stack

Open Command Palette (Cmd/Ctrl+Shift+P) → TokenLens AI: Select Stack Template

4. Send Your First Request

Select code in editor → Command Palette → TokenLens AI: Send Optimized Request

Model Routing

TokenLens AI automatically picks the right model for each task:

Task	Default Model	Why
Code completion	Claude Haiku	Fast, cheap, sufficient
Bug fix	Claude Sonnet	Balanced quality/cost
Architecture	Claude Opus	Complex reasoning needed
Documentation	GPT-4o-mini	Great writer, very cheap
Code review	Claude Sonnet	Balanced
Test generation	Claude Haiku	Pattern-based, cheap

Override any rule in Settings → tokenlens.router.*

Stack Templates

Pre-configured context profiles that know what matters for your framework:

Template	Includes	Excludes	Compresses
Next.js	types, lib, next.config	.next, node_modules	components, public
FastAPI	models, schemas, core	pycache, .venv	routers, tests
Django	models, serializers, config	migrations, pycache	views, tests

MCP Server Templates

Connect your tools in one click:

Service	Category	Features
🐙 GitHub	DevOps	PRs, issues, actions, code search
📋 Jira	Project Management	Issues, sprints, epics
🛒 Adobe Commerce	E-commerce	Products, orders, inventory, CMS
🐘 PostgreSQL	Database	Schema, queries, migrations
💬 Slack	Communication	Channels, threads, search
⚡ Linear	Project Management	Issues, cycles, roadmap
🐶 Datadog	Observability	Metrics, logs, traces, alerts
⚡ Supabase	Database + Auth	Schema, RLS, Edge Functions

Token Dashboard

The dashboard shows you exactly what happened with your last request:

👁 TokenLens AI

Session Usage
├── Total Tokens:    42,800
├── Tokens Saved:   180,200 (81%)
├── Total Cost:     $0.0184
└── Requests:       12

Last Request
├── Task Type:      bugfix
├── Model Used:     claude-sonnet-4-6
├── Raw Context:    38,400 tokens
├── After Optimize:  6,200 tokens  (84% saved)
├── Cost:           $0.0032
└── Duration:       1,240ms

Trimmed Files (12)
── node_modules/...
── .next/...
── prisma/migrations/...

Configuration

Full config via VS Code Settings or .tokenlens/config.json in your project:

{
  "providers": {
    "default": "anthropic",
    "anthropic": { "apiKey": "${ANTHROPIC_API_KEY}" },
    "openai":    { "apiKey": "${OPENAI_API_KEY}" },
    "ollama":    { "baseUrl": "http://localhost:11434" }
  },
  "budget": {
    "dailyTokenLimit": 500000,
    "alertAt": 0.8
  },
  "context": {
    "maxTokensPerRequest": 20000,
    "alwaysExclude": [".env", "node_modules", "*.lock"]
  },
  "stack": "nextjs",
  "mcp": {
    "github": { "enabled": true },
    "jira":   { "enabled": false }
  }
}

Contributing

We welcome contributions! The most wanted:

New MCP templates (Shopify, Salesforce, HubSpot, AWS...)
New stack templates (Go, Rust, Flutter, Laravel...)
Token optimization improvements
Bug reports and feature requests

See CONTRIBUTING.md for details.

Roadmap

[ ] LLM call inspector (DevTools for AI)
[ ] Inline eval runner
[ ] Multi-model benchmark per codebase
[ ] Background agents
[ ] Team token budget management
[ ] AI pipeline builder
[ ] VS Code fork (full editor)

License

MIT — see LICENSE

Built with ❤️ for AI developers. Star ⭐ if you find it useful!