catalyst-tracing 0.0.9

First-party OpenInference-shaped tracing for Python LLM and agent applications on Catalyst by Inference.net.

catalyst-tracing

First-party OpenInference-shaped tracing for Python LLM and agent applications running on Catalyst by Inference.net.

catalyst-tracing gives you one Python package for instrumenting common model SDKs, agent frameworks, and custom agent work. It emits OpenTelemetry spans with OpenInference-compatible attributes over OTLP/HTTP so Catalyst can display model calls, tool calls, prompts, responses, token usage, and parent-child agent flows.

This package is currently in beta. APIs may change before 1.0, but the package name and import path are intended to remain stable.

Install

Install the base tracing runtime:

pip install catalyst-tracing

Install only the integrations your application uses:

pip install 'catalyst-tracing[openai]'
pip install 'catalyst-tracing[anthropic]'
pip install 'catalyst-tracing[langchain]'
pip install 'catalyst-tracing[langgraph]'
pip install 'catalyst-tracing[langsmith]'
pip install 'catalyst-tracing[openai-agents]'
pip install 'catalyst-tracing[claude-agent-sdk]'
pip install 'catalyst-tracing[pydantic-ai]'
pip install 'catalyst-tracing[elevenlabs]'

You can combine extras. For LangChain agents that are already wrapped in LangSmith decorators, install both integrations plus the provider SDKs you use:

pip install 'catalyst-tracing[openai,anthropic,langchain]'
pip install 'catalyst-tracing[openai,langchain,langsmith]'

Quick Start

Set your Catalyst endpoint and token:

export CATALYST_OTLP_ENDPOINT="https://your-catalyst-otlp-endpoint"
export CATALYST_OTLP_TOKEN="your-token"
export CATALYST_SERVICE_NAME="checkout-agent"

Initialize tracing before creating SDK clients:

from catalyst_tracing import setup
from openai import OpenAI

tracing = setup()

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize this order."}],
)

tracing.shutdown()

The OpenAI call is captured as an OpenInference-shaped LLM span and exported to Catalyst through OTLP/HTTP.

What It Instruments

Integration	Install extra	What is captured
OpenAI	openai	Chat Completions, Responses, sync clients, and async clients
Anthropic	anthropic	Messages API calls, sync clients, and async clients
LangChain	langchain	Callback-manager driven chain, model, tool, and retriever spans
LangGraph	langgraph	Graph and node spans through the LangChain callback path
LangSmith	langsmith	LangSmith OpenTelemetry spans bridged into the Catalyst provider
OpenAI Agents	openai-agents	Agent runs plus nested OpenAI model spans
Claude Agent SDK	claude-agent-sdk	query() calls and yielded agent messages
Pydantic AI	pydantic-ai	Pydantic AI's native OpenTelemetry instrumentation
ElevenLabs Agents	elevenlabs	Conversation sessions, user/agent messages, and client tool calls

The base package includes the tracing runtime. Extras install the upstream SDKs themselves so you can keep production environments narrow.

LangSmith Decorator Integration

Applications that already use LangSmith decorators do not need Catalyst-specific wrappers around those functions. Keep using LangSmith's @traceable decorator and initialize Catalyst once at process startup. The SDK bridges LangSmith's OpenTelemetry context into the Catalyst tracer provider, so provider spans created below a decorated function are parented under that LangSmith run.

import os

from catalyst_tracing import setup
from langsmith import Client, traceable

os.environ["LANGSMITH_TRACING"] = "true"

tracing = setup(service_name="support-agent")
client = Client()

@traceable(name="lookup_order", run_type="tool", client=client, enabled=True)
def lookup_order(order_id: str) -> str:
    return f"order {order_id} is shipped"

lookup_order("ABC-123")
client.flush()
tracing.shutdown()

When LANGSMITH_TRACING=true is set and no LangSmith OTel mode is configured, Catalyst defaults LangSmith to hybrid mode. LangSmith still receives its own traces, and Catalyst receives the OpenTelemetry spans from the same decorators. Set LANGSMITH_TRACING_MODE=otel when you want LangSmith to emit only OTel spans for this process. The SDK also enables the LangSmith OTel context flags needed by older LangSmith clients, including LANGSMITH_OTEL_ENABLED, so the active decorator span becomes the parent for nested LangChain and provider spans.

The SDK enriches LangSmith spans before export so Catalyst can render them like the rest of your agent trace:

LangSmith signal	Catalyst/OpenInference attribute
run_type / langsmith.span.kind	openinference.span.kind
gen_ai.prompt	input.value with input.mime_type=application/json
gen_ai.completion	output.value with output.mime_type=application/json
tool trace name	tool.name
agent trace name	agent.name
LangSmith model metadata	llm.model_name, llm.provider

This also works with LangChain or LangGraph spans emitted under a LangSmith decorated root span. Catalyst captures LangChain chain/tool/retriever spans through its LangChain callback integration, captures provider spans through the provider integrations, and preserves the LangSmith decorator spans in the same trace tree.

For a LangChain agent with LangSmith tracing, the customer application should look like normal LangSmith and LangChain code:

import os

from catalyst_tracing import setup
from langchain.agents import create_agent
from langsmith import Client, traceable

os.environ["LANGSMITH_TRACING"] = "true"

tracing = setup(service_name="question-gen-agent")
langsmith_client = Client()

agent = create_agent(...)

@traceable(name="question_generation_agent", run_type="chain", client=langsmith_client)
def run_agent(payload: dict) -> dict:
    return agent.invoke(payload)

run_agent({"messages": [{"role": "user", "content": "Generate one question."}]})
langsmith_client.flush()
tracing.shutdown()

Do not add a separate customer-owned Catalyst wrapper around the decorated entry point just to make grouping work. If every model span appears as a root span, check that the process installed the langsmith extra, initialized Catalyst before invoking the agent, and did not override LangSmith's OTel mode after setup.

Public API

Most applications only need setup():

from catalyst_tracing import setup

tracing = setup(
    service_name="support-agent",
    service_version="0.4.0",
)

setup() returns a CatalystTracing handle with:

Attribute	Purpose
provider	OpenTelemetry TracerProvider configured for Catalyst export
tracer	Tracer for manual spans
install_results	Per-integration install results
shutdown()	Flush and close tracing before process exit

You can also import integration installers directly:

from catalyst_tracing import setup
from catalyst_tracing.openai import install_openai

tracing = setup()
install_openai(tracing.provider)

Available entry-point modules:

Import	Export
catalyst_tracing.openai	install_openai
catalyst_tracing.anthropic	install_anthropic
catalyst_tracing.langchain	install_langchain
catalyst_tracing.langgraph	install_langgraph
catalyst_tracing.langsmith	install_langsmith
catalyst_tracing.openai_agents	install_openai_agents
catalyst_tracing.claude_agent_sdk	install_claude_agent_sdk
catalyst_tracing.pydantic_ai	install_pydantic_ai
catalyst_tracing.elevenlabs	install_elevenlabs

Manual Agent Spans

Use manual_span() when work does not go through a supported SDK, such as a custom router, planner, evaluator, provider-failover step, or tool executor. It accepts OpenInference span kinds, provider-shaped usage payloads, structured inputs/outputs, and OTel-safe custom attributes.

from catalyst_tracing import SpanKindValues, manual_span, setup

tracing = setup()

with manual_span(
    tracing.tracer,
    name="question_generation/bloom",
    span_kind=SpanKindValues.CHAIN,
    system="fireworks",
    input={"template_id": "bloom", "total_questions": 10},
    model="accounts/fireworks/models/gpt-oss-120b",
    metadata={"deck_id": "deck_123"},
    tags=["question-gen-agent", "template:bloom"],
) as span:
    result = run_question_generation()
    span.set_output({"question_count": len(result.questions)})
    span.record_usage(result.usage)

tracing.shutdown()

agent_span() remains available as a focused convenience wrapper for AGENT spans:

from catalyst_tracing import agent_span, setup

tracing = setup()

with agent_span(tracing.tracer, name="RefundReviewAgent", system="internal") as span:
    span.set_input("Review refund request #1842")
    decision = run_refund_review()
    span.set_output(decision.summary)
    span.record_tokens(prompt=820, completion=160)

tracing.shutdown()

Any child spans created inside the context automatically parent under the agent span through standard OpenTelemetry context propagation.

When you already have an active span, reusable helpers are exported for common manual instrumentation tasks:

from catalyst_tracing import record_span_usage, set_span_attributes

record_span_usage(span, {"prompt_tokens": "820", "completion_tokens": 160})
set_span_attributes(span, {"metadata": {"deck_id": "deck_123"}})

Configuration

You can configure tracing with keyword arguments or environment variables.

Option	Environment variable	Default
endpoint	CATALYST_OTLP_ENDPOINT	http://localhost:8799
token	CATALYST_OTLP_TOKEN	unset
service_name	CATALYST_SERVICE_NAME	generated catalyst-app-* name
service_version	CATALYST_SERVICE_VERSION	0.0.6
debug	CATALYST_DEBUG	false
batching	none	"batch"

Legacy OTLP_ENDPOINT, OTLP_INGEST_TOKEN, and SERVICE_NAME variables are also accepted for compatibility.

Span Shape

Spans use OpenInference-style semantic attributes so LLM-aware viewers can understand them without custom adapters:

Attribute family	Examples
Span kind	openinference.span.kind
Inputs and outputs	input.value, output.value
Messages	llm.input_messages.*, llm.output_messages.*
Model metadata	llm.model_name, llm.invocation_parameters
Token counts	llm.token_count.prompt, llm.token_count.completion, llm.token_count.total
Provider/system	gen_ai.system

Constants are exported for custom spans:

from catalyst_tracing import Attr, SpanKindValues

span.set_attribute(Attr.SPAN_KIND, SpanKindValues.LLM.value)
span.set_attribute(Attr.MODEL_NAME, "gpt-4o-mini")

Error Handling

The package raises typed errors for misuse and returns structured install results for optional integrations:

from catalyst_tracing import CatalystTracingError, InvalidTracerProviderError
from catalyst_tracing.openai import install_openai

try:
    result = install_openai(provider)
except InvalidTracerProviderError as exc:
    print(exc.code)
except CatalystTracingError:
    raise

Each installer returns an InstrumentResult with:

Field	Meaning
name	Integration name
installed	Whether instrumentation was installed
code	Stable status code such as INSTALLED or SDK_NOT_INSTALLED
reason	Human-readable detail when installation is skipped

Package Names

The primary package is catalyst-tracing and the primary import path is catalyst_tracing.

Inference also publishes inference-catalyst-tracing as a company-qualified install name. It depends on this package and re-exports the same public API from the inference_catalyst_tracing import path.