AI framework integrations

Trace agent execution, tool calls, chains, and memory operations across LangChain, LlamaIndex, CrewAI, and 15+ more AI frameworks

OpenLIT instruments AI orchestration frameworks to give you end-to-end visibility into agentic workflows. Each span captures the framework operation type, inputs, outputs, and timing — so you can trace how a user request flows through chains, agents, and tool calls.

How framework tracing works

When you call openlit.init(), OpenLIT patches the framework’s internal execution methods. For each operation — a chain run, an agent step, a tool call, a retrieval — OpenLIT creates an OpenTelemetry span.

Setting detailed_tracing=True (the default) captures per-step traces for maximum observability. Set it to False to reduce span volume in high-throughput scenarios.

import openlit

# Detailed tracing enabled by default
openlit.init(otlp_endpoint="http://127.0.0.1:4318")

# Reduce span volume by disabling per-step traces
openlit.init(
    otlp_endpoint="http://127.0.0.1:4318",
    detailed_tracing=False
)

Get started

1. Install OpenLIT

   pip install openlit
   
2. Install your framework

   Install the framework package your application uses:

   # LangChain
   pip install langchain
   
   # CrewAI
   pip install crewai
   
   # LlamaIndex
   pip install llama-index
   
3. Initialize OpenLIT

   Call openlit.init() before any framework code runs.

   LangChain example

   import openlit
   from langchain_openai import ChatOpenAI
   from langchain_core.prompts import ChatPromptTemplate
   
   openlit.init(otlp_endpoint="http://127.0.0.1:4318")
   
   llm = ChatOpenAI(model="gpt-4o")
   prompt = ChatPromptTemplate.from_template("Tell me about {topic}")
   chain = prompt | llm
   
   result = chain.invoke({"topic": "observability"})
   print(result.content)
   

   import openlit
   from crewai import Agent, Task, Crew
   
   openlit.init(otlp_endpoint="http://127.0.0.1:4318")
   
   researcher = Agent(
       role="Researcher",
       goal="Research AI observability best practices",
       backstory="You are an expert in AI systems monitoring."
   )
   
   task = Task(
       description="Summarize the key benefits of AI observability",
       agent=researcher
   )
   
   crew = Crew(agents=[researcher], tasks=[task])
   result = crew.kickoff()
   print(result)
   

Supported frameworks

Framework	Python	TypeScript	Instrumentor key
LangChain	✅ langchain-core	✅	langchain
LangGraph	✅ langgraph	—	langgraph
LlamaIndex	✅ llama-index	✅	llama_index
CrewAI	✅ crewai	—	crewai
OpenAI Agents SDK	✅ openai-agents	—	openai-agents
Pydantic AI	✅ pydantic-ai	—	pydantic_ai
AG2 / AutoGen	✅ ag2 / pyautogen	—	ag2
Haystack	✅ haystack	—	haystack
Mem0	✅ mem0ai	—	mem0
Browser Use	✅ browser-use	—	browser-use
Letta	✅ letta-client	—	letta
Crawl4AI	✅ crawl4ai	—	crawl4ai
Firecrawl	✅ firecrawl	—	firecrawl
Dynamiq	✅ dynamiq	—	dynamiq
ControlFlow	✅ controlflow	—	controlflow
SwarmZero	✅ swarmzero	—	swarmzero
DSPy	✅ dspy	—	dspy
Agno (Phidata)	✅ agno	—	agno
GuardrailsAI	✅ guardrails-ai	—	guardrails

Span attributes captured

Framework spans include attributes specific to the operation type:

Attribute	Description
gen_ai.system	Framework name (e.g., langchain)
gen_ai.operation.name	Operation type (e.g., chain, agent, tool)
gen_ai.request.model	Underlying LLM model used
gen_ai.usage.input_tokens	Tokens consumed by the operation
gen_ai.usage.output_tokens	Tokens generated by the operation
db.cost	Estimated cost in USD
gen_ai.prompt	Input to the operation (when capture_message_content=True)
gen_ai.completion	Output from the operation (when capture_message_content=True)

Enable detailed_tracing=True (the default) to capture individual chain steps and agent iterations. This creates richer traces at the cost of additional span volume.

---

LLM provider integrations

Trace individual LLM calls with token counts, cost, and latency

Vector database integrations

Trace vector search, upsert, and index operations

Destinations

Send telemetry to Datadog, Grafana, New Relic, and other observability stacks