mikehasa/agentacct

See what your coding agents did and what it cost. Breaks each task down into work steps — tools used, files changed, tests run, time and tokens spent. Local-first dashboard for Claude Code, Codex, OpenCode, and more. No login, no telemetry.

What it solves

agentacct provides visibility into the activities of various coding agents (such as Claude Code, Codex, and Kimi Code) by acting as a local-first ledger. It solves the problem of not knowing exactly what an agent is doing, how much it costs, and whether its claims of completion are actually verified by tests or checks.

How it works

The tool reads local session logs from coding agents and processes two separate evidence streams:

  1. Usage Data: Extracts token counts and estimates costs based on pricing tables from the client's own session files.
  2. Activity Data: Captures tool calls, recorded steps, and machine checks (like test runs) via MCP (Model Context Protocol) or session logs.

It then joins these two streams using session and transcript IDs to create a "Work Receipt" that links specific actions to their resource consumption, assigning a confidence label to the accuracy of the attribution.

Who it’s for

For developers using multiple AI coding agents across different projects and clients, who need a transparent, local-only audit trail of agent activity, token usage, and cost estimates without sending data to the cloud.

Highlights

  • Local-First Privacy: All data stays on the user's machine; it does not request API keys or phone home.
  • Work Receipts: Combines activity timelines, recorded outcomes, and cost estimates into a single view.
  • Verification vs. Reporting: Distinguishes between an agent simply reporting a task as "done" and a result being "Verified" by passing checks.
  • Cross-Agent Tracking: Aggregates work across different agents and projects into shared timelines.
  • TUI and GUI: Offers both a terminal user interface and a native macOS app for monitoring usage and activity.
  • Non-Intrusive Recording: Records tool names, files touched, and exit codes, but explicitly ignores prompts and model responses to maintain privacy.

Related