AgentAcct

What it does

• Work Receipts join two evidence streams — what was used (tokens, pricing-table cost estimates) and what was done (steps and checks recorded over MCP) — linked on real session/transcript ids, every attribution carrying a confidence label (exact / high / medium / low)
• Reported is never Verified: an agent's own "done" needs current passing checks after the last recorded change, and failed attempts stay visible in the history alongside later results
• Usage analytics with independent date/agent/model/provider filters, fresh vs cache token splits, and provider quota windows — costs are ≈ pricing-table estimates, never invoices
• Work groups roll a project folder up across every agent on one cross-agent timeline, with partial coverage and overlap labeled rather than hidden
• Local-first by design: reads only local session files, listens on 127.0.0.1, stores commands credential-scrubbed — never your prompts, responses or transcripts

When to use it

Answering "what did my agents actually do all week?" across clients and projects; auditing before you trust an agent's "done"; comparing Claude Code vs Codex on the same task; spotting cost drift by model and day. There is also a paste-in prompt that lets your coding agent install and onboard it for you.

Getting started

pipx install agentacct
agentacct onboard   # once per machine: finds your agents, sets up a local store, starts the recorder
agentacct tui       # the terminal app

Requires Python >= 3.11 on macOS or Linux; Windows is supported only via WSL. A signed, notarized macOS app (macOS 14+) ships with the latest GitHub release. Open a new agent session afterwards — hooks and MCP servers bind at session start.

Things to know

Early alpha by its own words. Support is per capability, not per logo: Claude Code and Codex carry the fullest receipts today, while OpenCode, Hermes, DeepSeek Harness, Kimi Code, OpenClaw and Cursor are rated lane by lane in the coverage matrix (agentacct capabilities agents prints it from your installed version). Estimates are marked ≈ and low-confidence joins render as gaps, not zeros — it would rather show you a gap than a guess.