LangChain Integration

TraceAgent provides a native callback handler for LangChain that automatically captures chain steps, LLM calls, tool usage, and retriever activity.

Installation

pippip install trace-agent-sdk trace-agent-langchain

uvuv add trace-agent-sdk trace-agent-langchain

Basic Usage

from trace_agent_sdk import TraceAgentClient
from trace_agent_langchain import TraceAgentLangChainCallback

client = TraceAgentClient("http://localhost:8000")
run = client.start_run("langchain-agent", "Answer user question")

callback = TraceAgentLangChainCallback(run)
agent.invoke({"input": "What's the weather?"}, callbacks=[callback])

run.finish({"status": "success"})

Captured Events

Event	Captured Data
Chain start/end	Chain type, input, output, duration
LLM call/response	Prompt, response, token counts
Tool use/result	Tool name, parameters, output
Retriever query	Query, retrieved documents
Error	Error type, message, traceback

How Tools and Models Are Recorded

The callback reports each tool through a single call to the server, so a tool run appears once in the timeline with its real name and duration, and it is counted once in the run's tool and error totals. A failed tool counts as one error.

For every model call, the callback also remembers the provider and model, so the run shows them (and their token usage when the model reports it).

The callback never slows down or breaks your chain: if the TraceAgent server is down, the chain keeps running.

With Local Models

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="http://localhost:5000/v1",
    api_key="not-needed",
    model="local-model"
)

callback = TraceAgentLangChainCallback(run)
agent.invoke({"input": "..."}, callbacks=[callback])

tip

Set OPENAI_BASE_URL environment variable to avoid hardcoding the endpoint.

Environment Variables

Variable	Description
TRACE_AGENT_BASE_URL	Server URL for the callback handler
OPENAI_BASE_URL	OpenAI-compatible endpoint for local models
OPENAI_API_KEY	API key (use any value for local models)

What's Next?

• LangChain Example — Full walkthrough
• Pydantic AI Integration — The equivalent one-line integration for Pydantic AI agents
• OpenTelemetry Ingestion — Trace without TraceAgent code in your agent
• SDK Reference — Manual instrumentation