agentacct

Run three different coding agents against a project for a week and you end up with three different, hard-to-compare accounts of what actually happened, often with the agent's own chat transcript as the only record of what it changed, tested or spent. agentacct starts from a blunt version of that frustration - its own README literally asks 'what the fuck are my agents actually doing?' - and builds a local tool that answers it by reading the session logs agents already write to disk, rather than asking the agent to self-report after the fact.

It parses session files from Claude Code, Codex, OpenCode and Hermes and reconstructs them into 'work receipts': a per-task breakdown of tools invoked, files touched, tests run, time elapsed and tokens spent. Every claim in that receipt carries an evidence tier - reported (the agent said so), verified (a check actually passed) or observed (a hook recorded it happening) - so you can tell an agent's self-description apart from what was independently confirmed. A dashboard flags tasks that need review and tracks usage against provider limits, a terminal UI (`agentacct tui`) covers the same ground from the command line, and everything runs against a local API with no cloud round trip.

It fits anyone running more than one coding agent day to day and losing track of which one did what, especially teams that want a paper trail before trusting an agent's own summary of a task. The strict local-only design - no account, no cloud sync, no stored API keys - also suits developers who are simply not willing to send session data anywhere, whether for privacy reasons or because their employer would not allow it. Because it installs via pipx and ships a native macOS app on top of the CLI, it is approachable for solo developers as well as small teams comparing agent usage across a shared codebase.

The project is still young and effectively maintained by one primary author plus a close collaborator, a real bus-factor risk if development slows. Support for Hermes and OpenCode session logs is newer and less battle-tested than the Claude Code integration, and because each agent's log format is undocumented and can change without notice, agentacct's parsing has to keep chasing a moving target. It is MIT-licensed and fully OSI-open, but still pre-1.0 (v0.11.2), so the schema behind a 'work receipt' could still change between releases, worth checking before you build anything that depends on its JSON output.

The verdict: agentacct is worth adopting if you run multiple coding agents and want one verifiable, local record of what they did and what it cost, instead of trusting each agent's own summary. Its evidence-tier model, distinguishing reported claims from verified and observed ones, is a genuinely useful discipline that most agent tooling skips. Treat the multi-agent log parsing as still maturing, and the underlying data schema as not yet stable, so pin a version if you build automation on top of its output today.