agentacct vs codeburn
Local-first work receipts for coding agents: reads the session logs Claude Code, Codex, OpenCode and Hermes already write and reports what each task did, what it cost, and whether a check proved it. — versus — Local-first cost ledger for AI coding: reads the session files 36 tools already write and breaks every token and dollar down by task, model, project. TUI, web, desktop, menubar — no proxy, no keys.
Nearest neighbour and a clean split. Codeburn is a local-first cost ledger — tokens and dollars broken down by task, model and project from 36 tools' session files. Agentacct reads the same logs but reports the work and its evidence, and will not call a task verified without an independent check.
| agentacct | codeburn | |
|---|---|---|
| Stars | 667 | 9.7k |
| Forks | 75 | 769 |
| Language | Python | TypeScript |
| License | MIT | MIT |
| Last activity | today | yesterday |
| Topics | coding, local | coding, local |
| Curated connections | 7 | 3 |
agentacct — the curator's take
The only tool in this neighbourhood that refuses to take the agent's word for it. Everything else here counts tokens and dollars; agentacct counts work and evidence, and keeps them on separate axes — an agent saying 'done' files the task under Reported, and only a machine check that postdates the newest recorded work earns Verified. Every figure wears its basis (`≈` for estimated, bare ` only for reported), every field carries provenance (client hook, transcript scan, MCP record), and unconnected verifier sources stay honestly labelled 'not connected' instead of quietly scoring zero. macOS app, `agentacct tui`, and a loopback-only JSON API; no account, no telemetry, never touches your provider key. Use it when you are accountable for agent output. Overkill if you only want to know this month's spend.
codeburn — the curator's take
The honest architecture for spend tracking — read the session files already on disk, no proxy in your API path, nothing leaves the machine — executed at the broadest scope in the category: 36 tools, four surfaces, one data source. The parts beyond pretty dashboards earn the install: waste detection with applyable (and undoable) fixes, budget guards, model comparison. Limits are inherited from the source: it sees what session logs record, and subscription flat fees vs metered API blur what a 'dollar' means — treat per-task cost as a good estimate, not accounting (pricing via LiteLLM's daily refresh). When NOT: single-tool Claude Code forensics — cc-lens and claude-code-karma go deeper there with replay and cache breakdowns.