Skip to main content
On-prem privacyOpenAI · LLaMA Index · Open TelemetryJSON / JSONL · up to 10 MBNo SDK is required

Load Trace File

Supported formats: PDF (via GitHub), DOCX, ODT, EPUB, MOBI, HTML

LC
LangChain

JSON with run_id and nested child_runs

OT
OpenTelemetry

OTLP JSON with resourceSpans or flat JSONL with traceId+spanId.

LI
LlamaIndex

JSONL with event_type and parent_id (CBEventType).

How it works

Export the log

From LangChain use FileCallbackHandler. From LlamaIndex, CBEventHandler. From OpenTelemetry, otlp-json exporter.

Upload the file

JSON or JSONL up to 10 MB. The parser automatically detects the format.

Explore the tree

Expand nodes, click for details: Input, Output, Latency, Tokens, Cost €, Errors.

Privacy guaranteed

Everything happens in the browser. No data sent to the server. No required SDKs.

Unique Selling Proposition (USP) vs Langsmith/Heliconet

FunctionalityThis toolLangSmithHelicone
Required SDKNothing to translate!YesSi
Host in the CloudNeverAlwaysSempre
MultiformatLanguage Code + Operating System Type + License InformationLC onlyPartial
CostFreeFree and paid options.Free and paid options.

Come utilizzare AI Agent Trace Viewer

Export the trace from your framework

From LangChain use FileCallbackHandler or export with LangSmith; from LlamaIndex use CBEventHandler; from OpenTelemetry use otlp-json exporter: get a .json, .jsonl, or .ndjson file.

Drag and drop or upload the trace file

Use the drag-and-drop zone or click to select the file (max 10 MB). The parser automatically recognizes format: LangChain, LlamaIndex, or OpenTelemetry.

Explore the call tree

Expand or collapse nodes (LLM call, tool call, observation, decision): clicking on a node opens the details panel with input, output, latency, tokens, and estimated cost.

Analyze global statistics and errors

At top of results you'll find total latency, total tokens, estimated EUR cost, error rate, and number of nodes: useful for quickly identifying bottlenecks or failures in agents.

Suggerimenti

  • Use "Expand all" for a complete overview of the tree before focusing on nodes with high error rate or latency.
  • Mentally order analysis by bar-slow-latency-nodes (>8 seconds): these are prime candidates for performance optimization.
  • If working with large JSONL traces, ensure files don't exceed 5,000 rows: beyond that limit consider dividing them into multiple exports instead of uploading the whole file.

Domande frequenti

What trace formats does the viewer support?

It supports three formats: LangChain (JSON with run_id and child_runs nested), OpenTelemetry (OTLP JSON with resourceSpans or flat JSONL with traceId+spanId) and LlamaIndex (flat JSONL with event_type and parent_id from callback CBEventType).

Are my data sent to an external server or cloud?

No, parsing is entirely client-side: the file is read and processed in-browser via FileReader without any network calls. It's designed as a privacy-first alternative to cloud tools like LangSmith or Helicone.

What does estimated EUR cost per node represent?

It's an estimate of LLM call costs based on tokens registered in the trace, entering/exiting: if the trace lacks provider/model info or inference is on-premises, the cost may be €0.

Why has a 10 MB and 5,000 JSONL row limit been set?

It's a technical limitation to ensure fluid parsing and rendering within the browser since all processing occurs client-side without server-side streaming: for very large traces consider filtering or segmenting exports before loading.

Do I need to install an SDK to use this viewer?

No, unlike tools like LangSmith or Helicone you don't have to integrate any proprietary SDK into your codebase; just export the trace in standard formats supported by your framework (LangChain, LlamaIndex, OpenTelemetry) and upload here.