Agentacct

agentacct is a local-first ledger that shows what your coding agents have been doing and what it cost. It reads the local session files of clients such as Claude Code, Codex, Kimi Code, OpenCode, and Hermes, and presents recent activity, tool calls, recorded steps, checks, tokens, and estimated cost across projects. There is no account and no cloud: it reads only local files, listens only on the loopback interface, and never stores or requests a provider API key.

The core is a split between two evidence streams. What an agent used comes from each client's own session files, with tokens as the client reported them and costs as pricing-table estimates. What it did comes from steps and checks the agent records over MCP (Model Context Protocol) as it works, plus machine checks such as test runs. The two are joined by session and transcript id, and every attribution carries a confidence label, so an unproven link shows as a gap rather than a zero.

It separates what an agent claims from what is verified: a reported result is the agent's own done, while a verified result needs current passing checks after the last recorded change. A terminal app and an optional native macOS window present the same data. Views include per-task Work Receipts, folder-anchored work groups that span every agent, a dashboard of recent activity, and a usage page with independent date, agent, model, and provider filters that separate fresh, cache, and total tokens.

It records the work rather than the conversation: tool names and categories, touched files, credential-scrubbed commands, exit codes, tokens, and steps, never prompts, model responses, or transcripts. Support is rated per capability in a coverage matrix, with Claude Code and Codex carrying the fullest receipts.