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

• 🔀 Multi-model cost comparison — token count + estimated cost across all your configured models, side by side. This is the whole point.
• 🛈 Hover tooltip — hover over a selection or a string literal (JS/TS/Python, plus Markdown/JSON/YAML/plaintext) to see the comparison inline.
• 🔎 CodeLens on prompts — a token-count + cost lens appears right above detected prompt strings in your code. Click it for the full breakdown.
• 📊 Live status bar — select any text and its token count + cost appear in the status bar instantly (no command needed); the tooltip shows every model.
• 🗔 Comparison Panel — a visual dashboard comparing the whole catalog with a live output-token slider (recomputes instantly), sortable columns, and a "configured models only" filter.
• ⌨️ Commands — Estimate Selection, Estimate Clipboard, Open Comparison Panel, and Select Models to Compare.
• 🗂 25+ models — GPT-4o/4.1/o-series, Claude 4/3.7/3.5, Gemini 2.5/2.0/1.5, DeepSeek, Mistral, Llama, Grok. Pick what you care about.
• 📁 Workspace scan — one command finds every prompt in your project and reports the total estimated cost per run, with a clickable per-file breakdown.
• ➕ Custom models — add your own models/prices (or a negotiated rate) in settings; no need to edit bundled files.
• ⚠️ Context-window warnings — a ⚠ appears when a prompt exceeds a model's context window.
• 📋 Copy as Markdown — drop a ready-to-paste comparison table on your clipboard.
• 🔒 Local-first, zero config — exact OpenAI tokenization runs entirely on your machine; pricing is bundled. No network calls, no API key.

A note on accuracy

• OpenAI models use real BPE tokenization (o200k_base for the GPT-4o family, cl100k_base for GPT-4 / GPT-3.5) via js-tiktoken.
• Anthropic, Google, and everyone else (DeepSeek, Mistral, Llama, Grok, …) 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.)

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

This repo also ships an MCP server that exposes the same tokenizer + pricing engine as tools (estimate_cost, count_tokens, list_models) to Claude Desktop, Claude Code, Cursor, or any MCP client — so you can ask "what does this prompt cost on GPT-4o vs Claude vs Gemini?" right in your chat. See mcp-server/README.md for setup. Same local-first, no-API-key core — just a different front end.

Install

From the Marketplace (once published):

1. Open the Extensions view (Ctrl+Shift+X / Cmd+Shift+X).
2. Search for "LLM Cost & Token Estimator".
3. Click Install.

From a .vsix:

code --install-extension llm-cost-estimator-0.3.2.vsix

From source (for development): see Contributing.

Usage

• Hover: hover over a string literal — or select text — in a supported file. A tooltip shows the per-model token count and cost.
• CodeLens: open a JS/TS/Python file with prompt strings; a N tokens · ~$X · compare lens sits above each one. Click it for the full breakdown.
• Live status bar: select any text — the status bar instantly shows its token count and cheapest cost; hover the item for the full comparison. Click it to open the Comparison Panel.
• Comparison Panel: Command Palette → LLM Cost: Open Comparison Panel (or the editor toolbar icon). Drag the output-token slider to see costs update live, sort by any column, or filter to your configured models.
• Commands (Command Palette, Ctrl+Shift+P):
  • LLM Cost: Estimate Selection — selection, or the whole file if nothing is selected. Also on the editor right-click menu.
  • LLM Cost: Estimate Clipboard — estimate whatever you've copied.
  • LLM Cost: Open Comparison Panel — the visual dashboard.
  • LLM Cost: Scan Workspace for Prompts — project-wide prompt cost report.
  • LLM Cost: Copy Comparison as Markdown — table to clipboard.
  • LLM Cost: Select Models to Compare — pick models from the catalog.
  • LLM Cost: Refresh Pricing from the Web — pull current prices (see below).
  • LLM Cost: Reset Session Total — clear the running total.

How "total" is calculated: cost = input tokens + an assumed number of output tokens (output pricing is usually higher than input, so it matters). The assumption is configurable and always shown in the tooltip.

cost = (inputTokens  / 1e6) * inputPer1M
     + (outputTokens / 1e6) * outputPer1M

Settings

All settings live under llmCostEstimator.*:

Available model keys (out of the box) — run LLM Cost: Select Models to Compare to pick from these visually:

• OpenAI: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3, o4-mini, gpt-4-turbo, gpt-4, gpt-3.5-turbo
• Anthropic: claude-opus, claude-sonnet, claude-haiku, claude-3.7-sonnet, claude-3.5-sonnet, claude-3-opus
• Google: gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash, gemini-1.5-pro, gemini-1.5-flash
• Others: deepseek-chat, deepseek-reasoner, mistral-large, mistral-small, llama-3.3-70b, llama-3.1-405b, grok-2

Example settings.json:

{
  "llmCostEstimator.models": ["gpt-4o", "gpt-4o-mini", "claude-sonnet", "gemini-1.5-flash"],
  "llmCostEstimator.outputTokenAssumption": 800,
  "llmCostEstimator.currency": "USD"
}

Custom models

Add your own models — or override a built-in price with a negotiated rate — without touching the bundled files, via llmCostEstimator.customModels:

{
  "llmCostEstimator.customModels": {
    "my-finetune": {
      "label": "My Fine-tune",
      "provider": "openai", // "openai" = exact tokenization; anything else = estimate
      "inputPer1M": 1.0,
      "outputPer1M": 2.0,
      "contextWindow": 128000 // optional, enables the ⚠ over-limit warning
    },
    "gpt-4o": { "label": "GPT-4o (our rate)", "provider": "openai", "inputPer1M": 2.0, "outputPer1M": 8.0 }
  }
}

Then add the key to llmCostEstimator.models (or pick it via Select Models to Compare). Invalid entries are reported and skipped, not silently dropped.

Live pricing

The bundled prices are representative placeholders. Run LLM Cost: Refresh Pricing from the Web (or enable llmCostEstimator.refreshPricingOnStartup) to pull current prices from a maintained source — by default LiteLLM's pricing data. The result is cached, so it keeps working offline and falls back to bundled prices on error.

Each model in pricing.json declares a liveId (the source's key for that model); models without one keep their bundled price. Point llmCostEstimator.pricingSourceUrl at any LiteLLM-compatible JSON to use a different source. User customModels always win over live prices.

Updating pricing (bundled defaults)

⚠️ The bundled prices are representative placeholders and change frequently. Verify them against each provider's official pricing page before relying on them.

• OpenAI — https://openai.com/api/pricing/
• Anthropic — https://www.anthropic.com/pricing
• Google — https://ai.google.dev/pricing

Prices live in src/pricing/pricing.json, keyed by model alias. Each entry looks like:

"gpt-4o": {
  "label": "GPT-4o",
  "provider": "openai",
  "encoding": "o200k_base",
  "inputPer1M": 2.5,
  "outputPer1M": 10.0
}

• inputPer1M / outputPer1M are USD per 1,000,000 tokens.
• provider is openai, anthropic, or google (determines the tokenizer and whether the count is exact or an estimate).
• encoding is the BPE used to count tokens — o200k_base or cl100k_base. For Anthropic/Google it's only an approximation proxy.

To add a model or change a price, edit the JSON and rebuild (npm run compile). The pricing file is bundled into the extension, so changes take effect after a rebuild/reinstall. Pull requests that keep prices current are very welcome.

Contributing

Contributions are welcome — bug reports, pricing updates, new providers, and features alike. New here? Start with CONTRIBUTING.md, browse the ROADMAP, or pick a good first issue. By participating you agree to our Code of Conduct.

git clone https://github.com/waqarulwahab/llm-cost-estimator.git
cd llm-cost-estimator
npm install

npm test           # run unit + load tests (Vitest)
npm run test:load  # just the load/performance suite
npm run e2e        # bundle + end-to-end test against a mocked VS Code
npm run lint       # ESLint
npm run typecheck  # tsc --noEmit
npm run compile    # bundle to dist/ with esbuild

Then press F5 in VS Code to launch the Extension Development Host and try your changes live.

Project layout:

src/
  tokenizer/   # Tokenizer interface + per-provider implementations
  pricing/     # pricing.json + lookup & cost math
  core/        # estimator, prompt detector, workspace scan, export, formatting (all pure)
  ui/          # hover, status bar, CodeLens, QuickPick, comparison + scan webviews
  commands/    # command handlers
  extension.ts # activate() / deactivate()
test/          # Vitest unit + load tests for the core logic
mcp-server/    # MCP server (reuses src/core, src/pricing) — use it in Claude/Cursor

The core/, tokenizer/, and pricing/ layers are intentionally free of any vscode import so they can be unit-tested directly.

This project is licensed under the MIT License.

Packaging & publishing

The extension is bundled with esbuild and packaged with @vscode/vsce; the MCP server is published to npm. Both are automated via GitHub Actions on a version tag.

npm run package       # production bundle -> dist/extension.js
npm run vsce:package  # create the .vsix
npm run mcp:build     # build the MCP server
npm run mcp:test      # build + stdio end-to-end test the MCP server

👉 Full step-by-step guide — Marketplace, Open VSX, npm, GitHub Actions, secrets, and how end users install each artifact — is in PUBLISHING.md.

---

Built with js-tiktoken. Not affiliated with OpenAI, Anthropic, or Google.