Most coding-agent cost reports end at a total: tokens by model, dollars by day, perhaps a session count. That is useful bookkeeping, but it does not tell a developer whether a costly run changed the right files, ran a test, or merely spent a long time circling an error.
agentacct is an early-alpha, local-first project that tries to connect those questions. Its README says it imports usage from local client session files while recording task steps and machine checks, then joins the two evidence streams using real session or transcript identifiers.
The design detail I find most valuable is not the dashboard. It is the vocabulary around uncertainty:
- Imported token usage is described as
client_reported. - Costs are pricing-table estimates, not provider invoices.
- A passing machine check is distinct from an agent's own claim.
- Attribution can be exact, high, medium, or low confidence; an unprovable join can remain a gap.
That is a healthier model for agent engineering. A low-confidence attribution should not silently become a per-task cost fact just because the UI needs a number. Likewise, a statement that an agent “ran tests” should not carry the same weight as a recorded passing test.
There is an operational trade-off. The project documents Python 3.11+ and onboarding that may create global or project-scoped state and add client integration configuration. “Local-first” can be a privacy advantage, but it is not the same as zero review. The repository also calls the project early alpha, so a developer should read the install runbook and usage truth table before enabling it on a work machine.
I have not tested or run agentacct. This is a read-only review of its public README, documentation, repository metadata, and source layout; compatibility, performance, and integration behavior need independent verification. The project is MIT licensed according to its license file.
Top comments (0)