agentacct

agentacct is a local-first "Agent Work Intelligence" tool that answers a practical question: what are my coding agents actually doing, and what did it cost? It reads the local session logs of coding agents and presents recent activity, tool calls, recorded steps, checks, tokens and estimated cost across projects and clients. Supported clients mentioned include Claude Code, Codex, Kimi Code, OpenCode, Hermes and others, with a documented coverage matrix rating each capability lane separately. The README notes Claude Code and Codex currently carry the fullest receipts. Installation is via pipx (Python >= 3.11 on macOS or Linux; Windows only through WSL), with a signed macOS app also offered. Setup runs `agentacct onboard` once per machine to detect agents, create a local store and start a recorder; `agentacct tui` opens a terminal app. A project-scoped install keeps everything inside the repository. The interface centers on a Sessions view and a "Work Receipt" per task, combining participating sessions, recorded outcome, estimated cost, claims and check results, plus an activity timeline. A key distinction is Reported versus Verified: an agent's own "done" is only Reported, while Verified requires current passing checks after the last recorded change. Failed attempts remain visible in history. Work groups let you point at a project folder and see which agents worked there, with session details and a shared cross-agent timeline. The Dashboard shows latest recorded work with client, time, outcome, evidence and cost, and keeps historical failed checks and blockers in a separate "Recorded issues" area rather than making them the default ordering. Usage views filter independently by date, agent, model and provider, with range totals by agent and model, a daily chart with Fresh/All token toggles, and breakdowns separating cache reads and writes. Tokens come from client records; costs are pricing-table estimates marked with ≈ and are explicitly not invoices. Provider-reported quota windows and reset times appear on the same page. The terminal UI offers receipts, usage, capacity and a review overview, including a folder-anchored Work tab with a cross-agent timeline navigable by keyboard. The project emphasizes honesty and privacy: estimates are labeled, missing data is shown as a gap rather than a zero, every join between usage and recorded work carries a confidence label (exact, high, medium, low), and support is described per capability rather than per logo. It records work metadata such as tool names and categories, touched files, credential-scrubbed commands, exit codes, tokens and recorded steps, but not prompts, model responses or transcripts. It reads only local session files, listens only on 127.0.0.1, and never stores or requests provider API keys. Architecturally it keeps two evidence streams separate: what was used (from client session files) and what was done (from steps and checks recorded over MCP plus machine checks such as test runs), joined on real client ids with labeled confidence. Claude Code binds ids through an installed hook bridge; Codex and OpenCode are matched from their own logs at import time; Hermes joins only on explicitly passed ids. Uninstall steps stop managed processes, remove autostart if installed, uninstall the package, and optionally delete the local store and client config entries. Documentation covers reference material, coverage matrix, architecture, safety boundaries, privacy threat model, worked examples and a full demo.