AI Card Bao NAS.AICARDBAO.COM

Industry News

LiteLLM Lens collects Claude Code, Codex, OpenCode, and Pi calls into the same trace, and can also play back the entire coding session on the terminal

When someone in the team uses Claude Code, someone uses Codex, and someone uses OpenCode at the same time, the most troublesome question for managers is: where is the money spent, who spends it, and what is gained in exchange. Each tool has its own log format, which cannot be combined together.

The two functions LiteLLM has added to Lens in the past two days are for this purpose.

The first is unified collection. Model calls issued by the four coding agents, Claude Code, Codex, OpenCode, and Pi, can now be imported into Lens as tracking data and attributed according to the three dimensions of key, user, and team. How many tokens someone burned on which project can be seen in a table.

The collected conversations can also be played back, and the interface is made to look like a terminal. When something goes wrong, you don’t need to go through a colleague’s local record. You can just play that paragraph over again to see what the agent adjusted and what it got at each step.

The second one is Lens CLI, which is used to answer another question: How different is the result of the same task if the model or agent framework is changed. It supports cross-comparison of models and frameworks in the terminal. On the framework side, you can fill in Claude Agent SDK, OpenAI Agents SDK, LangGraph, or your own agent. Output four items: pass rate, cost, delay, and failure classification.

Failure classification is a more practical one. If you just look at the pass rate, you can't tell the difference between 60% and 70%; only by breaking it down by failure type can you know whether there are too many errors in tool calls, or whether the model simply understands the task wrongly.

LiteLLM is originally the gateway between applications and various models. All calls pass through it, making tracking easy. It should be noted that this set of things assumes that your encoding agent traffic actually passes through it. The part directly connected to the official interface cannot be seen by Lens.

Related