Cheap per call, expensive per hour

Why the spend is invisible until it is large#

Per call, an agent is trivially cheap — small enough that nobody checks. Per hour, with several agents each re-reading files, retrying, and carrying long context, it is not. Neither number is hidden, exactly; they are just kept somewhere you never look, in units that do not add up to a question you would ask. "Was this week expensive?" is not answerable from a list of API calls.

The provider dashboard does not close the gap either. It knows what your account spent; it does not know which agent, which repository, or which session spent it. When the bill is higher than expected, that is precisely the breakdown you need — and precisely the one nobody has.

One ledger, on your machine#

Because every model call already passes through a local proxy, it can be written down where it happens. Mirasim keeps one ledger of every call — tokens, cache hits, estimated cost — decoded and computed on your machine and nowhere else. Pick a window (today, 7, 30 or 90 days) and it leads with the numbers that answer the question directly.

• Calls, total tokens, output tokens, cache hit rate and estimated cost — each shown against the previous period, so "more than usual" is a fact rather than a feeling.
• The same window broken down by agent, model, account, workspace and session. This is the part a provider dashboard cannot give you.
• How much of the work went through the quota fallback instead of your own subscription. If that share is high, the bottleneck is your plan, not your prompts.

What it changes in practice#

The useful outcome is not frugality, it is attribution. Once spend is broken down by model and by session, the expensive habits become visible and specific: one long-context loop that re-reads the same tree every turn; a reasoning level left high on a task that never needed it; a cache hit rate that dropped because a prompt now changes every call.

It also makes model choice an empirical decision rather than a vibe. Sessions carry their own model and reasoning level, so the cheaper option is one switch away — and after a week you can look at the ledger and see whether the cheaper one actually cost you anything in results.