Token Meter

Your meter for
Claude Code & Codex.

Local-first dashboard that shows how many tokens you spent, where they went, and how fast each model and MCP tool actually runs. Your JSONL never leaves your machine.

Free + MIT core, available now. Pro ($5/mo) is live — drop your email for Pro+ & new releases.

npm install -g @whdrnr2583/token-meter
token-meter ingest && token-meter serve

See it in one command

Ask your AI assistant "this week's spend" over MCP, or run the CLI. Same data, zero cloud.

$ token-meter stats 7

 Day   Model              Calls      USD   %day
 ───────────────────────────────────────────────
 Mon   claude-opus-4-7      372   $41.20    99%
       claude-haiku-4-5      45    $0.34     1%
 Tue   claude-opus-4-7      394   $44.10   100%
 Wed   claude-opus-4-7      521   $39.80    99%
       claude-sonnet-4-6    114    $0.41     1%
 ───────────────────────────────────────────────
 Total 7 days             4,257  $312.40   100%

 Heaviest : Read 1.1M · Bash 731k · WebSearch 451k  (resp tokens)
 Slowest  : Agent 123s avg (237x) — your latency sink
 Cache    : 94% reuse — on a flat-fee plan, not billed per token

Why Token Meter

📊 Multi-vendor in one view

Claude Code ~/.claude/projects/ and Codex ~/.codex/sessions/ merged. No SDK, no proxy.

🔍 MCP & tool breakdown

See exactly which MCP server eats your tokens and which tool is slow. Latency, response size, call count.

💵 USD-equivalent cost

Your Max plan is a flat fee — but what would the API have cost? Token Meter shows the saving (or burn). Estimates only — not your actual invoice.

🏠 Local-first, private

SQLite on your own machine. No upload, no cloud account required for the free tier.

⚡ MIT licensed core

The CLI and dashboard are open source. Fork it, audit it, ship it.

🕒 Time-of-day insight

Find your most expensive hours, your slowest periods, and the days that blew up the budget.

🔔 Smart alerts

Threshold rules with desktop and webhook actions. Pipe into Slack, Discord, n8n — your call.

Connect to your AI tool

Token Meter ships an MCP server — once registered, your AI tool can ask "this week's spend", "recent sessions to resume", "why was session X expensive" without you opening the dashboard.

🚀 Easiest: one command

npx -y @whdrnr2583/token-meter install-mcp all

Registers Claude Code + Cursor + Claude Desktop in one go. Idempotent, backs up existing config, preserves your other MCP servers. Add --dry-run to preview.

🤖 Or let your LLM handle it

Open Claude Code / Cursor / Claude Desktop and paste:

Read https://raw.githubusercontent.com/whdrnr2583-cmd/token-meter/main/docs/mcp-server.md and set up Token Meter as my MCP server.

Or set it up by hand

Claude Code

claude mcp add token-meter -- \
  npx -y @whdrnr2583/token-meter mcp

Verify with claude mcp list.

Cursor

Edit ~/.cursor/mcp.json (Windows: %USERPROFILE%\.cursor\mcp.json):

{
  "mcpServers": {
    "token-meter": {
      "command": "npx",
      "args": ["-y", "@whdrnr2583/token-meter", "mcp"]
    }
  }
}

Restart Cursor → Settings → MCP shows green.

Claude Desktop

Edit claude_desktop_config.json with the same JSON as Cursor.

Path: ~/Library/Application Support/Claude/ (mac), %APPDATA%\Claude\ (Windows). Fully quit and reopen.

ChatGPT / Other

ChatGPT requires HTTP — wrap with mcpo (community recipe). Direct stdio works with Continue, Zed, any MCP client:

npx -y @whdrnr2583/token-meter mcp

Full guide with verification + troubleshooting: docs/mcp-server.md

Pricing

Free

$0

  • Claude Code + Codex parsing
  • MCP & tool breakdown
  • Project / model / hour breakdown
  • 7-day history
  • USD-equivalent cost
  • 1 desktop alert rule

Pro live

$5/mo

  • Everything in Free
  • 30-day history
  • Smart alerts unlimited — desktop + webhook
  • Session drill-down — find the exact session that blew the budget
  • Cache efficiency — hit ratio + how much caching saved you
  • Waste signals — oversized tool responses + cache that didn't pay off

After payment we email your license key (tm_live_...) to your Polar checkout email. Activate with token-meter activate <key>. Full Pro spec →

Pro+ later

~$24/mo

  • Unlimited history
  • Local LLM proxy (Ollama, LM Studio, llama.cpp, vLLM)
  • GPU / VRAM tracking — millisecond TTFT/ITL
  • Behavior-changing automation (MCP auto-trim, model switching, cloud↔local routing)
  • Model benchmark lab
  • Multi-machine sync (encrypted)
  • PDF reports (auto-emailed)

Substantial engineering — ships when the proxy + per-OS GPU work is scheduled in. Full Pro+ spec →