If you're running Claude Code or Codex locally and want to know where your tokens are actually going, this parses the JSONL files they already write to disk and breaks it down by model, project, MCP server, and individual tool. The MCP server exposes usage_summary, recent_sessions, and tool_stats so you can query cost and latency from inside your AI editor without switching to a dashboard. The free tier gives you 7 days of history and per-tool breakdowns. Everything stays local in a SQLite database under ~/.tokenpulse. Useful when you need to spot which MCP integration is burning tokens or compare hourly patterns across projects without sending your session data anywhere.
claude mcp add --transport stdio whdrnr2583-cmd-token-meter -- npx -y @whdrnr2583/token-meter