Respan Python SDK

respan.ai | Documentation | PyPI

respan-ai is the unified Python entry point for Respan tracing, decorators, context propagation, and first-party instrumentation plugins. It automatically discovers eligible direct LLM SDK adapters while keeping frameworks, agents, wrappers, and observability bridges explicit to avoid duplicate spans.

Requirements

• Python 3.11, 3.12, or 3.13
• A Respan API key for telemetry export
• The provider SDK used by your application

Installation

Install the facade and your provider SDK. For OpenAI:

pip install respan-ai openai

With Poetry:

poetry add respan-ai openai

Replace openai with the provider SDK your application uses. respan-ai already bundles the supported first-party instrumentation adapters and respan-tracing; provider SDKs remain application-owned dependencies.

Automatic Onboarding

Set your Respan credentials:

export RESPAN_API_KEY="your-api-key-here"
# Optional; this is also the Respan gateway base URL used below.
export RESPAN_BASE_URL="https://api.respan.ai/api"

Construct Respan before the first provider call. Python discovery and activation happen synchronously in the constructor, so there is no separate initialize() call:

import os

from openai import OpenAI
from respan import Respan

# Omit `instrumentations` to use curated direct-LLM automatic discovery.
respan = Respan()

client = OpenAI(
    api_key=os.environ["RESPAN_API_KEY"],
    base_url=os.getenv("RESPAN_BASE_URL", "https://api.respan.ai/api"),
)

response = client.chat.completions.create(
    model="gpt-4.1-nano",
    messages=[{"role": "user", "content": "Say hello in three languages."}],
)
print(response.choices[0].message.content)

print(respan.get_auto_instrumentation_status())
respan.shutdown()

This example routes inference through the Respan gateway. When calling a provider directly, configure that provider's credentials and endpoint normally; Respan still reads RESPAN_API_KEY for telemetry export.

Automatically Supported Providers

The facade bundles these adapters. Install only the provider SDKs your application actually imports.

Only registry entries classified as direct LLM integrations and enabled by default are activated. A missing provider SDK is reported as missing and does not stop application startup. Broad OpenTelemetry entry-point discovery remains disabled.

Inspect Instrumentation Status

for entry in respan.get_auto_instrumentation_status():
    print(entry["id"], entry["status"], entry.get("reason"))

Each entry includes the registry ID, runtime name, provider, provider SDK, instrumentation package, status, and optional reason. Status is one of enabled, disabled, missing, or failed.

The dataclass form is also available as respan.auto_instrumentation_status.

Control Automatic Discovery

Disable automatic discovery completely:

from respan import Respan

respan = Respan(is_auto_instrument=False)

Supplying an explicit instrumentation list selects explicit mode and disables automatic discovery by default:

pip install respan-ai respan-instrumentation-openai-agents openai-agents

from respan import Respan
from respan_instrumentation_openai_agents import OpenAIAgentsInstrumentor

respan = Respan(
    instrumentations=[OpenAIAgentsInstrumentor()],
)

To deliberately combine an explicit plugin with direct-SDK automatic discovery:

respan = Respan(
    instrumentations=[OpenAIAgentsInstrumentor()],
    is_auto_instrument=True,
)

Explicit plugins activate first, so a matching automatic adapter is skipped rather than activated twice. Python currently provides a global automatic-discovery switch, but not a per-provider disable option.

Tracing Helpers

The facade re-exports Respan's tracing helpers:

from respan import agent, respan_span_attributes, task, tool, workflow

Use @workflow, @task, @agent, and @tool to structure traces, and use respan_span_attributes or propagate_attributes to attach customer, session, thread, environment, and metadata values.

Lifecycle

• Respan() initializes telemetry and activates eligible adapters.
• respan.flush() exports buffered spans without deactivating adapters.
• respan.shutdown() deactivates adapters and flushes telemetry.

Examples

• OpenAI SDK examples
• OpenAI Agents SDK examples
• CrewAI example

Public API

The package exports Respan, the automatic instrumentation registry and status types, OTELInstrumentor, the instrumentation protocol, tracing decorators, RespanClient, get_client, respan_span_attributes, and propagate_attributes.

License

Apache 2.0 — see the repository LICENSE.

Support

• Email: team@respan.ai
• Documentation: https://www.respan.ai/docs
• Issues: GitHub Issues