Hi HN, I'm the author. I kept using AI coding agents (Claude Code, Copilot) with no idea what any single prompt was actually costing me — I'd only find out from a bill or a quota wall later. So I built Toketer: a tiny status-bar meter that shows token usage and cost per prompt, live, right in the editor.
A few honest notes on how it works, because precision varies by tool:
- Claude Code: tracked exactly and automatically from its local logs (real token counts).
- GitHub Copilot: no extension can read Copilot's live token usage (there's no public API for it), so Toketer shows a clearly-labelled estimate (an "~" and an "est" badge, with a token range). I never present a guess as an exact number. You can also import your real Copilot usage for exact figures.
Privacy is the core design constraint: Toketer is metadata-only. It never reads or sends your code, files, or prompts anywhere — only token counts, model labels, and costs. It works offline, needs zero config, and is free and open source (MIT).
I'd genuinely love feedback — especially on the Copilot estimation approach and what other tools/agents you'd want supported. Happy to answer anything.
Hi HN, I'm the author. I kept using AI coding agents (Claude Code, Copilot) with no idea what any single prompt was actually costing me — I'd only find out from a bill or a quota wall later. So I built Toketer: a tiny status-bar meter that shows token usage and cost per prompt, live, right in the editor.
A few honest notes on how it works, because precision varies by tool:
- Claude Code: tracked exactly and automatically from its local logs (real token counts). - GitHub Copilot: no extension can read Copilot's live token usage (there's no public API for it), so Toketer shows a clearly-labelled estimate (an "~" and an "est" badge, with a token range). I never present a guess as an exact number. You can also import your real Copilot usage for exact figures.
Privacy is the core design constraint: Toketer is metadata-only. It never reads or sends your code, files, or prompts anywhere — only token counts, model labels, and costs. It works offline, needs zero config, and is free and open source (MIT).
It runs in both VS Code and Cursor (also on Open VSX: https://open-vsx.org/extension/toketer/toketer). GitHub repo: https://github.com/Soumyadeep-ThinkCode/toketer.git
I'd genuinely love feedback — especially on the Copilot estimation approach and what other tools/agents you'd want supported. Happy to answer anything.