Metadata-Version: 2.5
Name: havrin
Version: 0.2.0
Summary: See where your Claude Code usage goes, from your own local session logs
Project-URL: Homepage, https://havrin.dev
Author: Gavin Schiller
License-Expression: MIT
License-File: havrin/LICENSE
Keywords: claude,claude-code,cost,llm,usage
Classifier: Environment :: Console
Classifier: Intended Audience :: Developers
Classifier: Programming Language :: Python :: 3
Requires-Python: >=3.10
Description-Content-Type: text/markdown

# havrin

See where your Claude Code usage goes, from the session logs Claude Code already keeps on your machine.

```sh
pipx install havrin        # or: pip install havrin
havrin meter --days 30 --what-if --subagent-types
```

The meter reads the JSONL transcripts in `~/.claude/projects` and prints:

- tokens by kind (input, cache writes, cache reads, output) and each kind's cost-weighted share;
- which tools filled the context, in characters;
- with `--what-if`: the share of cost you would have saved with subagents on Sonnet, with the main context capped
  (150k tokens by default), and with both. These are estimates replayed from your own logs, not promises.
- with `--subagent-types`: subagent cost broken down by subagent type, and whether the Agent call asked for a model;
- with `--context-ranking`: the kinds of tool output (`git diff`, `pytest`, whole-file reads...) that fill the
  context most, and how much went to re-reading files that hadn't changed.

**Privacy:** it only reads local files, sends nothing anywhere, and prints counts, never the content of your sessions.

Options: `--days N` (default 7), `--sessions` (one line per transcript), `--dir PATH`, `--sonnet-ratio`, `--cap`.
Python 3.10+, no dependencies.

Havrin is building the rest: routing routine Claude Code steps to cheaper models and your own machine, with your
tests checking every result. Join the beta at [havrin.dev](https://havrin.dev).
