tokviz

See where your LLM tokens go.

One command. Real-time token usage and cost dashboard. No database, no Docker, no config.

npx tokviz

tokviz receives standard OpenTelemetry traces, filters for LLM spans (gen_ai.* attributes), and displays token consumption and costs in a live dashboard. When you stop it, it's gone. Nothing persists.

Why

You're building an app that calls LLMs. You want to know:

• How many tokens each request burns
• What it costs per interaction
• Which models are actually being used (routing, fallbacks)
• Whether something is silently eating your budget

Existing tools are either too heavy (Langfuse, OpenLIT — databases, Docker, SDKs) or too generic (Aspire Dashboard — shows everything, highlights nothing). tokviz shows you LLM spend and nothing else.

Quick start

npx tokviz

Point your OTLP exporter at http://localhost:4318 and you're done.

Works with Dev Proxy, the OpenTelemetry SDKs, and anything that emits gen_ai.* span attributes following the OpenTelemetry Gen AI semantic conventions.

With Dev Proxy

{
  "plugins": [{
    "name": "OpenAITelemetryPlugin",
    "enabled": true,
    "pluginPath": "~appFolder/plugins/dev-proxy-plugins.dll"
  }]
}

With OpenTelemetry SDK

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318

What you see

Overview — Token usage and cost over time, summary cards for total tokens, cost, requests, and unique models.

Models — Breakdown per response model: request count, input/output tokens, total cost.

Requests — Every request with timestamp, model, operation, tokens, cost, and duration. Click any row to see all span attributes.

Options

npx tokviz [options]

--port <number>   Port to listen on (default: 4318)
--no-open         Don't auto-open the browser

How it works

tokviz is an OTLP HTTP receiver that:

1. Accepts traces on POST /v1/traces (protobuf and JSON)
2. Filters spans — only processes those with gen_ai.* attributes
3. Stores everything in memory — no database, no disk writes
4. Pushes updates to the browser over WebSocket — zero polling

Supports both application/x-protobuf and application/json content types.

Development

git clone https://github.com/waldekmastykarz/tokviz.git
cd tokviz
npm install
npm run dev

Opens the Vite dev server at http://localhost:5173 with hot reload. The Express backend runs on port 4318.

License

MIT