Codebuff is a multi-agent terminal coding tool: you give it one task, and instead of answering itself, it dispatches a set of specialized sub-agents — a planner, a file editor, a reviewer, and more — that each talk to the model on your behalf. That design is great for getting work done hands-off. It is terrible for keeping a mental tally of spend, because the token cost of a single prompt can be scattered across many model calls. The collector does not claim to reconstruct sub-agent roles or parent tasks; it reports the day-and-model totals the saved records can support.

How do I see my Codebuff token usage?

One command reads the usage history Codebuff already writes locally and produces a readable daily and model total:

$ npx whoburnedmore↳ scanning local Codebuff logs…↳ Codebuff usage history detected   CODEBUFF USAGE  ────────────────────────────────────────  2026-06-16  input  4,180,300  output  402,700  2026-06-15  input  2,905,100  output  271,400  2026-06-14  input  3,640,900  output  338,200   BY MODEL  recorded-model-a      9.4M tokens  recorded-model-b      5.9M tokens  other models          3.1M tokens   7-day total:  18.4M tokens   est. cost shown per model

The figures above illustrate the report shape rather than literal CLI output. whoburnedmore walks the local entries and adds input and output token fields into daily and model buckets. It does not promise a per-request, per-role, or parent-task total when that attribution is not present in the accounting source.

What the local logs actually contain

The useful accounting fields are model identifiers, input and output token counts, and dates. whoburnedmore uses that numeric usage data; prompts, code, and file names are never included in a leaderboard submission.

1. 1

   Read the recorded usage rows

   The collector reads the source's accounting output and keeps only the fields needed for token totals. It does not infer sub-agent roles or task lineage.
2. 2

   Split by model

   Different recorded models can have very different public API rates. Keeping their usage separate makes the estimate more useful.
3. 3

   Roll up by day

   Daily totals surface the heavy days — the ones where a single ambitious task fanned out into a long chain of agent calls — instead of burying them in a monthly average.

Why is Codebuff's multi-agent cost hard to track?

The fan-out is the whole story. One thing you type becomes many model calls, and each sub-agent often re-sends shared context so it can do its job — which multiplies input tokens in a way that is invisible from the outside 🧮:

Because a multi-agent run can create many calls, a daily total may be much larger than the number of prompts you typed. What the local accounting can safely show is the model mix, not which internal role produced each token:

How much is Codebuff costing me?

Estimated cost is the sum of each recorded model bucket, with input and output tokens priced separately. whoburnedmore applies published API rates and adds the buckets:

The reason this matters: a task that felt cheap because you only typed one sentence can be expensive once you count the model calls it caused. Pricing varies by tool and plan, so whoburnedmore does not guess your plan — it estimates the underlying API cost of the tokens you actually spent, which is the honest yardstick whether you are on credits or a flat fee.

One tool or all of them

If Codebuff is the only agentic tool you run, the totals are just Codebuff. If you also use Claude Code, Codex, or Gemini CLI, the same command shows them side by side — see the cross-tool usage guide for that combined view.

Is reading my Codebuff logs private?

Prompts, code, and file names are never submitted. The default command does submit daily aggregate usage rows after sign-in. Run npx whoburnedmore --dry-run to print exactly what a submission would contain, or npx whoburnedmore --local to open a dashboard that stays entirely on your computer. 🛡️