agentlens-sdk

agentlens-sdk

Python SDK for AgentLens — real-time observability and debugging for AI agents.

Install

pip install agentlens-sdk

With framework extras:

pip install "agentlens-sdk[openai]"
pip install "agentlens-sdk[anthropic]"
pip install "agentlens-sdk[langchain]"
pip install "agentlens-sdk[all]"

Prerequisites

Start the AgentLens server first:

pip install agentlens-server
agentlens-server
# Server running at http://localhost:8766

Then open the dashboard at http://localhost:5173.

Usage

Option 1: Auto-instrument (recommended)

from agentlens_sdk import auto_instrument

auto_instrument()  # patches OpenAI + Anthropic clients automatically

# All subsequent OpenAI/Anthropic calls are traced — no other changes needed
import openai
client = openai.AsyncOpenAI()
response = await client.chat.completions.create(...)

Option 2: @trace decorator

from agentlens_sdk import init, trace

init(server_url="ws://localhost:8766/ws")

@trace(name="my_agent")
async def my_agent(query: str) -> str:
    result = await call_llm(query)
    return result

Option 3: Manual spans

from agentlens_sdk import get_tracer

async def my_agent(query: str) -> str:
    with get_tracer().span("my_agent") as span:
        span.set_attribute("query", query)
        result = await call_llm(query)
        span.set_output(result)
        return result

Framework Integrations

LangChain

from agentlens_sdk.interceptors.langchain_interceptor import AgentLensCallbackHandler

handler = AgentLensCallbackHandler()
chain = my_chain.with_config(callbacks=[handler])

CrewAI

from agentlens_sdk import auto_instrument
auto_instrument()  # CrewAI is auto-detected

AutoGen

from agentlens_sdk import auto_instrument
auto_instrument()  # AutoGen ConversableAgent is auto-detected

Configuration

from agentlens_sdk import init

init(
    server_url="ws://localhost:8766/ws",   # AgentLens server WebSocket URL
    http_url="http://localhost:8766",       # AgentLens server HTTP URL
    session_name="my-agent-v2",            # Optional: name shown in dashboard
)

Design Principles

• Never raises exceptions — all SDK operations are silent on failure. Your agent always runs.
• Non-blocking — async, buffered, batch-sent. Adds <5ms overhead per traced call.
• Auto session management — creates a session on first event, closes it on process exit.
• Sensitive data redaction — fields named api_key, token, password, secret, authorization are automatically replaced with [REDACTED].
• Graceful degradation — if the server isn't running, events are silently dropped.

License

MIT