Get 2,500 events tracked for freeSign up now

Who it is for

The prompt change that doubled token use shipped three weeks ago

You are iterating on prompts, tools and graphs, and every one of those changes moves cost. The feedback loop on that is the monthly invoice, which is slow enough that the causal link is gone by the time you see the number — and the invoice cannot tell you which agent or which change was responsible anyway.

What a provider dashboard leaves you with

  • 01

    System prompts assembled by interpolation defeat prompt caching, silently. We found 347 such call sites across 39 repositories.

  • 02

    Frontier models left in test paths bill at frontier rates on every run.

  • 03

    Calls behind a framework are the hardest to attribute, and 62% of the repositories we scanned make them that way.

What Capsera does about it

Cost per run, visible immediately
Attribution at the run level, live, so a prompt change's cost shows up in the same session you made it.
Framework instrumentation
LangGraph, CrewAI, LangChain, LlamaIndex and AutoGen are instrumented, so a graph composed at runtime is attributable without manual tagging.
Thirteen providers
Sync and async clients across model APIs, inference platforms and cloud model services, patched automatically at init.

The number this is judged on

Cost per run, per agent version

None of this is assumed. We scanned 133 public agent repositories and found a cost defect in most of the ones making live model calls.

Read the method and limitations

See what that prompt change cost, today.

Stop runaway agents before they become runaway invoices.

Try for free