LLM Cost Estimator
Estimate multi-model API cost from tokens or pasted prompts.
OpenTurn agent trace JSON into a readable waterfall of spans, durations and LLM attributes.
Use an OTLP, Jaeger or OpenInference JSON export.
Spans appear in start order with timing bars.
Model, token and prompt attributes show under each span.
Review the summary JSON for sensitive text, then share it.
An agent run took forty seconds and you want to know whether the time went to the model, a tool call or a retry loop. The trace export is a few thousand lines of nested JSON. Agent Trace Studio turns it into a timeline you can read.
Paste trace JSON and click Visualize. The format is detected automatically: OTLP (resourceSpans with scopeSpans or the older instrumentationLibrarySpans), Jaeger's data array with processes for service names, or an OpenInference-style spans array, including a bare array of spans. The waterfall lists spans by start time, each with its duration in milliseconds, service name, a short span id and a bar positioned on the overall timeline. Under each span, up to six LLM-related attributes are shown: keys starting with gen_ai., llm., openinference., ai. or langchain, and keys mentioning token, prompt, completion or model.
The summary panel holds JSON with the detected format, span count, how many spans carry LLM attributes, and each span's name, duration, service and those attributes. Check it before sharing: attributes whose keys mention prompt or completion can contain the actual text. If they do, mask that text on your device first. The token attributes in a trace pair well with the LLM Cost Estimator when you want to price a run.
Parsing happens in your browser, and the Receipt button records a local note of the session, as explained in how privacy receipts work.
Keyboard shortcut
Answers for this tool. For site-wide help, open the FAQ hub.
For OpenInference and flat span arrays, the tool subtracts start from end exactly as the values appear in the file. If your export uses nanosecond timestamps, the bars stay in proportion but the millisecond labels will be scaled.
Not necessarily. The banner adds up every span duration, and child spans overlap their parents, so the sum is usually larger than the wall-clock time. Read the root span duration for the end-to-end figure.
There is no filter control. Spans are listed in start order with their durations, so long bars stand out visually, and you can search the summary JSON in your editor for a specific name.