How it works / How we parse usage
How we parse usage
What the CLI reads from your agent logs — and what it ignores.
Centrail's CLI reads the usage logs your coding agent already writes — today,
Claude Code (~/.claude/projects), GitHub Copilot CLI
(~/.copilot/session-state), and Codex (~/.codex/sessions).
What a usage event is
Each event is one model request. From it we read:
| Field | Example |
|---|---|
| Model | claude-opus-4-8 |
| Tokens | input, output, cache read, cache write |
| Timestamp | when the request happened |
| Working dir | which repo it ran in |
What we never read
Your source code. Your prompts. Your completions. The parser only extracts the counts and metadata above — never content.
One model, one maker. A model is always counted toward the company that built it —
gpt-5.3-codexcounts as OpenAI even when you reach it through Copilot. Copilot is multi-provider, so the same connector also brings Claude and Gemini usage, each attributed to its maker. A standalone Gemini connector is coming.