AgentCost Tracks Local Agent Spend

This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.

AgentCost 0.1.0 reads local coding-agent transcripts and produces usage reports. The package documents adapters for Claude Code, Cursor and Codex, plus Ollama recording. Its dollar figures are estimates, not an alternative billing ledger.

This assessment uses the published package documentation. We have not independently validated its accounting against provider invoices. The practical question is whether a report can explain an expensive session without turning incomplete logs into false precision.

Inspect the data before comparing costs

The documented report includes a data-quality field and pricing date. Some records can be estimated or placeholders. A total that looks exact can therefore mix different kinds of evidence.

Run the documented inspection commands in an environment where you have approved access to the relevant transcripts:

agentcost doctor
agentcost providers
agentcost list --provider cursor
agentcost pricing --check

The package name is agentcost-cli; its command is agentcost. Check the installed version’s help before choosing a report or session selector. Avoid assuming that an example from a newer release works in 0.1.0.

Local transcript paths can contain prompts, file names and code. Keep reports with the same access restrictions as the underlying logs, and remove private content before attaching one to a public issue. “Local” describes where processing happens; it does not make the input nonsensitive.

Reconcile one session before ranking tools

Choose one completed, narrow task whose diff you can understand. Record the session identifier, tool version, model and approximate start and end times. Match the report to that task before interpreting any dollar figure.

Use this review checklist:

If the provider exposes usage for the same interval, compare it with the report. Explain differences such as cached tokens, model routing, missing records or subscription billing before calling the local estimate accurate. Do not infer Cursor quota consumption directly from a third-party dollar estimate.

Use discrepancies to improve the task boundary

A report can suggest repeated context loading or a long tool loop. It does not establish why the work took that path. Read the corresponding transcript segment and diff before rewriting repository instructions.

For example, a task limited to one billing helper may still require reading callers elsewhere. That can be necessary investigation. Repeatedly reopening the same unrelated file without new evidence is a different problem. The receipt helps locate the behavior; review determines whether it was waste.

For a small exercise, compare two repetitions of the same kind of task, preserving both accepted and rejected results. Record review time alongside token usage. Keep the experiment local and descriptive until enough comparable runs exist to support a broader cost claim.

Updated 21 September 2026: removed repeated keyword phrases, added documented inspection commands, and separated usage estimates from invoices and product limits.