Hermes Agent is Nous Research's self-improving terminal coding agent — the one with persistent memory across runs and a sandboxed execution loop. Its headline feature is reach: it can drive 300+ models across a long list of providers, picking whichever fits the step in front of it. Great for results, rough for accounting. A model-level local history is useful because it shows which recorded models drove the total without pretending to reconcile every provider invoice.

How do I see my Hermes Agent token usage?

Hermes keeps local usage records that the collector can turn into daily and model totals. The default command includes a sign-in because it publishes aggregate rows; choose--local before running if you only want the on-device dashboard:

$ npx whoburnedmore↳ reading Hermes Agent run history…↳ Hermes usage history detected   HERMES AGENT — TOKENS BY DAY  ────────────────────────────────────────  2026-06-17  in 4,118,300  out 312,700  2026-06-16  in 2,640,900  out 201,400  2026-06-15  in 5,902,100  out 470,800   BY MODEL  recorded-model-a    11.4M tokens  recorded-model-b     6.8M tokens  recorded-model-c     2.1M tokens   7-day total:  31.7M tokens   est. $128.90

The figures above demonstrate the shape of the report rather than literal CLI output. whoburnedmore reads the history off disk and attaches a published API rate when it can identify the model. The result is an API-equivalent estimate, not a provider-invoice reconciliation and not a subscription charge.

How do I compare Hermes model usage?

This is the part that makes Hermes different from a single-model tool. When one run touches several models, the saved model buckets show which ones carried the most token volume. They do not prove which provider billed a call, so the honest visualization is a model mix 📊:

Token share and cost share are not the same chart. A smaller token bucket can contribute more estimated cost when its model has a higher published rate. That comparison can inform model-routing decisions without inventing task roles or provider-level attribution.

Why model variety hides your real number

A single grand total hides whether one premium model or many cheaper models drove the week. Reading the local day-and-model records keeps those buckets separate. You can then compare the estimate with provider billing, while treating the provider dashboard as authoritative for the amount actually charged.

How much is Hermes Agent costing me?

Cost is just usage times a rate — but with Hermes the rate is not one number, it is a different price for every model recorded. So the estimate is a sum over model buckets, priced with the public rate table:

We will not quote per-token prices here, because for Hermes there is no single price to quote — model rates move. What matters is the method: whoburnedmore reads each recorded model bucket, looks up a published API rate, and adds it in. This provides a consistent comparison; your provider's billing page remains authoritative for discounts, credits, and routing.

Where does Hermes keep this data, and is it private?

Hermes writes its run history to a local directory on the machine that ran the agent — alongside the persistent memory it uses to improve between sessions. whoburnedmore reads numeric usage fields such as token counts, model names, and dates. Prompts, code, and file names are never submitted. 🛡️

$ npx whoburnedmore --dry-runwould submit (aggregates only):  tokens_in_total   31,742,800  tokens_out_total   2,544,100  models_seen        9  prompts / code / filenames:  never submitted

Use --dry-run to print exactly what a submission would contain, or npx whoburnedmore --local to keep the whole breakdown on your own machine and skip the leaderboard entirely. Either way, the agent's memory and your source stay where Hermes left them.