Open-source Python SDKs for tracebility, the self-hosted LLM observability + eval-rigor + agent-replay platform.
| Package | Install | What it is |
|---|---|---|
tracebility |
pip install tracebility |
Native SDK. The canonical surface — IngestClient + ControlClient, @trace decorator, with span() context manager. Use this for new code. |
tracebility-langsmith |
pip install tracebility-langsmith |
Drop-in compat shim for the langsmith package. One-line import migration: from langsmith import Client, traceable → from tracebility_langsmith import Client, traceable. Same call shapes, same decorators. |
tracebility-langchain |
pip install tracebility-langchain |
LangChain / LangGraph callback bridge. TracebilityCallbackHandler you pass to Runnable.invoke(..., config={"callbacks": [...]}). Tree-aware, thread-safe. |
All three are independent: install only what you need.
pip install tracebilityfrom tracebility import TracebilityClient, trace
client = TracebilityClient(
endpoint="https://your-tracebility.example.com",
api_key="tb_...",
)
@trace
def my_agent(prompt: str) -> str:
# your agent logic
return llm_call(prompt)Java, Go, Rust, .NET, Ruby, etc. should send traces via OpenTelemetry directly:
OTEL_EXPORTER_OTLP_ENDPOINT=https://your-tracebility.example.com/v1/tracesThe tracebility ingest accepts OTLP HTTP/JSON and translates OpenInference + OTel GenAI semantic conventions to native runs/spans automatically. No SDK required in those languages.
| Source | What to install |
|---|---|
| New code | tracebility |
| Existing LangSmith code | tracebility-langsmith (one-line import change) |
| LangChain / LangGraph apps | tracebility-langchain (drop in a callback handler) |
| Other Python frameworks | tracebility + manual instrumentation |
Apache-2.0. See LICENSE.
PRs welcome. Each package has its own pyproject.toml; develop locally with:
cd tracebility # or tracebility-langsmith / tracebility-langchain
pip install -e ".[dev]"
pytest