Splunk open-sources a tool that prices AI coding sessions from local trace files

Install it on Monday and the first thing on screen is last week. The agents wrote their trace files as they went, so the reader has a cost history before anyone starts a new session ^[2]. Nothing has to be instrumented, and nothing has to be shipped off the box. devops.com's platform line reads "Linux, macOS and Linux", so devops.com does not say whether Windows developers are covered ^[4].

Because the tool sits where the agent sits, the threshold notification reaches the person who ran the expensive session ^[9][1]. Whether that helps depends on which problem you have. It tells a developer who had no idea what a two-hour agent run costs; it does not hand a finance owner one number for the whole team.

The columns worth opening first are the ones that explain why a session cost what it did. Token Meter splits fresh input from generated output, reports how long the run spent waiting on the model, and tracks context growth, retries and failures across the run ^[5][6]. A session that got expensive because the agent kept re-sending a swelling context needs a different fix from one that got expensive because it wrote a lot of code.

Hao Yang, vice president and head of AI for Splunk, said DevOps teams should be able to strike a better balance between consuming tokens and relying on platforms from, for example, Splunk to minimize token consumption, according to devops.com ^[11]. Yang also said that how agents access and process data will become much more federated as model providers look for ways to help organizations keep costs under control ^[12]. The free tool tells you what you spent, and Splunk sells the platform that makes the number smaller.

The framing figure in the devops.com write-up is a Futurum Group survey finding that just under half of organizations are over their AI budgets ^[14]. That also means a slim majority are inside them ^[16]. That figure is measured per organization, while Token Meter measures per machine.

If you want to know which model should have run a given task, the tool answers that per session and per model, and one developer can start with it tomorrow ^[7][18]. If you want to know what the team spent last month, the data is on laptops you do not administer, and the version devops.com describes reads local logs one machine at a time ^[1]. The second question needs somewhere to put the logs before any of this is useful for it.