Agentmeter

Context

AI coding sessions leave useful usage data behind, but the logs are local, provider-specific, and awkward to compare over time. Agentmeter turns Claude Code and Codex session logs into a private usage history without introducing a hosted telemetry service.

One local data path

The scan command reads the standard session directories, normalizes events, and stores them in a versioned SQLite database. serve then exposes a private browser dashboard with token and estimated-cost history, while the CLI can also emit machine-readable JSON for other tools.

The first working version reconciles its totals against independent reference calculations. That makes the dashboard more than a log viewer: imported usage is checked before it becomes a number a user might rely on.

Storage and pricing correctness

The database uses transactional migrations and idempotent batch insertion, so rescanning the same logs does not duplicate usage events. Unknown models remain visible, but their costs are marked incomplete rather than silently treated as free. A bundled pricing snapshot handles known models while keeping the accounting honest when a model is not recognized.

Privacy boundary

Session logs can contain source code, file contents, and prompts. Agentmeter keeps the data path local and explicitly forbids real session data in the repository; test fixtures are synthetic and hand-authored. That boundary is part of the product rather than an afterthought in the README.

Verification and delivery

The repository’s make check gate runs formatting, go vet, linting, and race-enabled tests. The same project can produce CGO-free binaries for Windows and Linux, keeping deployment as a single executable with no runtime service to install.

Outcome

Agentmeter has moved beyond scaffolding into a working development version with scanning, local reporting, SQLite persistence, and a browser dashboard. The repository remains actively developed as the supported session formats and reporting surface mature.