agentacct 0.9.0 — four agents, one Work Receipt

The multi-agent release. agentacct now reads and instruments four coding agents — Claude Code, Codex, OpenCode, and Hermes — and turns each task into one honest Work Receipt that works across all of them. Ships alongside the first signed, notarized macOS app.

Highlights

• Four coding agents, one Work Receipt. Usage, cost, sessions, and now WHAT each session did (its commands, edited files, and tools) — from each agent's own on-disk store, projected through one client-agnostic Task and Receipt model. agentacct onboard instruments each installed agent it finds.
• Receipt Actions derived from a client's own transcript — no hook required. For agents whose hooks don't fire (Codex, OpenCode), the commands, edited files, and tools are recovered at import time from the store agentacct already scans for tokens. Only names, cwd-relative paths, and scrubbed commands are derived — never tool output.
• Independent checks for OpenCode, from its recorded exit codes: a recognized test/build/lint run lifts a step from self-checked to independently-checked — the one local signal that is not the agent's own word.
• Honest Actions provenance — the Receipt names whether each session's Actions came from a live client hook or a scan of the client's own store; nothing scan-derived is labelled as hook-observed.
• The macOS app. Download agentacct-0.9.0.dmg below, drag agentacct to Applications, and open it — on first launch it installs the bundled CLI, instruments the agents it finds, and shows your Work Receipts in a native window. Signed + notarized; requires macOS 14+.

Install

• CLI: pipx install agentacct (or uv tool install agentacct), then agentacct onboard.
• macOS app: the signed, notarized .dmg attached below.

Everything stays on your machine — no account, no cloud sync, no phone-home. See the full CHANGELOG.