LLM Cost Estimator MCP server

LLM Cost & Token Estimator

See how many tokens your prompt uses — and what it'll cost on GPT-4o vs. Claude vs. Gemini — right inside VS Code. No API key, no billing dashboard, no guessing.

Why this exists

When you're building an LLM-powered app, two questions come up constantly: how many tokens is this prompt? and what will this call cost? Today you either guess, paste into a web tokenizer, or check the provider's billing console after the fact — all of which pull you out of your editor.

This extension answers both questions inline, and it answers them for several models at once, so you can make a real price/quality trade-off ("~$X on GPT-4o, ~$Y on Claude Sonnet, ~$Z on Claude Haiku") without leaving the file you're working in. It's local-first and works with zero configuration — no API key required.

Demo

Hover any prompt (or select text) to see its token count and cost across your

models, side by side — cheapest first. Same comparison powers the CodeLens, the

live status bar, and the Comparison Panel. (animated version)

Features

your configured models, side by side. This is the whole point.

• 🔀 Multi-model cost comparison — token count + estimated cost across all

plus Markdown/JSON/YAML/plaintext) to see the comparison inline.

• 🛈 Hover tooltip — hover over a selection or a string literal (JS/TS/Python,

detected prompt strings in your code. Click it for the full breakdown.

• 🔎 CodeLens on prompts — a token-count + cost lens appears right above

the status bar instantly (no command needed); the tooltip shows every model.

• 📊 Live status bar — select any text and its token count + cost appear in

with a live output-token slider (recomputes instantly), sortable columns, and a "configured models only" filter.

• 🗔 Comparison Panel — a visual dashboard comparing the whole catalog

Panel, and Select Models to Compare.

• ⌨️ Commands — Estimate Selection, Estimate Clipboard, Open Comparison

DeepSeek, Mistral, Llama, Grok. Pick what you care about.

• 🗂 25+ models — GPT-4o/4.1/o-series, Claude 4/3.7/3.5, Gemini 2.5/2.0/1.5,

reports the total estimated cost per run, with a clickable per-file breakdown.

• 📁 Workspace scan — one command finds every prompt in your project and

settings; no need to edit bundled files.

• ➕ Custom models — add your own models/prices (or a negotiated rate) in

context window.

• ⚠️ Context-window warnings — a ⚠ appears when a prompt exceeds a model's

clipboard.

• 📋 Copy as Markdown — drop a ready-to-paste comparison table on your

your machine; pricing is bundled. No network calls, no API key.

• 🔒 Local-first, zero config — exact OpenAI tokenization runs entirely on

A note on accuracy

family, cl100k_base for GPT-4 / GPT-3.5) via js-tiktoken.

• OpenAI models use real BPE tokenization (o200k_base for the GPT-4o

do not publish reliable local tokenizers, so their counts are approximated using an OpenAI encoding and are clearly marked with a ~ and a disclaimer in the UI. They're great for ballpark cost comparison, not for exact billing. (An optional API-based accurate mode is a candidate for a future release.)

• Anthropic, Google, and everyone else (DeepSeek, Mistral, Llama, Grok, …)

Also available as an MCP server (use it in Claude / Cursor)

Shortened. The full README is on GitHub.