Show HN: Frugal Tokens – explore costs and usage across coding agents

Frugal Tokens, a developer tool for tracking and analyzing spend across AI coding agent sessions, launched on Hacker News on August 19, 2026, per a Show HN post by author dpc94.

The tool visualizes overall token usage, estimated working time, overlapping sessions, and a breakdown of spend by model and cache misses, according to the Show HN post. At the session level, it provides percentile breakdowns and a list of sessions with high-level metadata. Clicking into any session opens an explorer showing individual model calls alongside tool inputs and outputs, including the ability to jump directly to where a cache miss occurred. It also includes a rough cost comparison showing what a recorded session would have cost under a different model's pricing, or — for Anthropic users — under 5-minute versus 1-hour caching configurations.

The project originated from personal curiosity, as dpc94 explained in the Show HN post: "I wanted to share a project I've been working on called Frugal Tokens. I originally built it because I was curious to see how much all of my sessions cost and how much cache misses affected that spend." The author noted observing "widely different spend profiles" among users and wanted to understand what drives those differences. The tool requires Deno and is available as a single command to run once Deno is installed; source code is published on GitHub. A live demo with some data scrubbed is also available.

The release comes as developers using AI coding agents — tools like Claude, Copilot, and similar LLM-backed assistants — report highly variable and sometimes opaque costs. Cache miss behavior in particular can significantly inflate token spend, but most agent interfaces offer little visibility into where those misses occur or how session structure affects them.

Looking ahead, the author said in the Show HN post they want to collect more data to identify which workflow patterns drive higher costs, specifically calling out long sessions, high context usage, and many turns as areas of interest. No company, funding, or team beyond the solo author has been disclosed.