Itemized receipts for your coding agents

What it does

agentacct is a local-first dashboard that ingests session logs already written by coding agents such as Claude Code and Codex. It correlates token usage with actual work steps—tests run, files changed, tools used—and surfaces the breakdown in a browser bound to your machine alone. State lives in plain local files, the UI stays on 127.0.0.1, and the tool never stores provider API keys or phones home.

The interesting bit

Instead of trusting an agent’s self-reported summary, agentacct joins client usage logs with work records captured over MCP, then labels every correlation with an attribution confidence score. When it cannot prove a link, it renders the gap rather than guessing. Costs are similarly honest: local pricing-table estimates clearly marked as such, never invoice truth.

Key highlights

• Reads existing local session files from Claude Code, Codex, and others without touching your repositories
• Correlates token burn to specific tasks and labels every join exact, high, medium, or low
• Runs fully offline with plain-file state, a localhost dashboard, and no accounts or telemetry
• Surfaces machine checks like passing tests as Verified evidence while keeping agent claims tagged Agent reported
• Early alpha with a published per-client capability matrix; some integrations are narrower than others

Caveats

• Windows is only supported via WSL; macOS and Linux are first-class
• Cost figures are estimates from a local pricing table, not actual provider invoices
• Attribution coverage varies by client: Claude Code and Codex have the strongest join mechanics, while others like Cursor or OpenCode have more limited paths

Verdict

Developers already running Claude Code or Codex who want to audit their token spend will find this useful. If you do not use local coding agents or you need cloud-hosted enterprise analytics, this is not your tool yet.