bryel 0.1.1

OpenTelemetry-native tracing for LLM apps and agents — ships OpenInference/OTLP traces to bryel.

bryel (Python)

OpenTelemetry-native tracing for LLM apps and agents. One call ships your model calls, tool calls, tokens, and cost to bryel as OpenInference traces — from any framework.

uv add bryel            # or: pip install bryel

Quickstart

import bryel
bryel.init(api_key="bk_…", service_name="my-app")

# Trace OpenAI (any OpenInference instrumentor works the same way):
from openinference.instrumentation.openai import OpenAIInstrumentor
OpenAIInstrumentor().instrument()

from openai import OpenAI
OpenAI().chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "hello"}],
)

Prefer bryel to wire up everything it can find? Let it instrument every installed OpenInference library for you:

import bryel
bryel.init(api_key="bk_…", instrument=True)   # instruments openai, anthropic, langchain, …

How it works

bryel is just the exporter. It configures an OpenTelemetry TracerProvider with an OTLP/HTTP exporter pointed at the bryel ingest, authenticated with your API key. The OpenInference instrumentors emit the spans; bryel's ingest normalizes both the OpenInference (llm.*) and OpenTelemetry GenAI (gen_ai.*) conventions — so there's nothing to map.

API

bryel.init(api_key=None, *, service_name=None, endpoint=None, headers=None, set_global=True, instrument=False)

Configures the exporter and returns the TracerProvider.

• api_key — your project API key (bk_…). Falls back to $BRYEL_KEY.
• service_name — OpenTelemetry service.name.
• endpoint — OTLP/HTTP traces URL. Defaults to the bryel ingest, or $BRYEL_INGEST_URL for self-hosting.
• headers — extra headers, merged after auth.
• set_global — register as the global OTel provider (default True).
• instrument — also auto-instrument installed OpenInference libraries.

bryel.auto_instrument(tracer_provider=None) -> list[str]

Best-effort instrument every installed OpenInference library — provider SDKs (OpenAI, Anthropic, Bedrock, Groq, Mistral, VertexAI, Google GenAI, LiteLLM, Instructor), agent SDKs (OpenAI Agents, Google ADK), and frameworks (LangChain/LangGraph, LlamaIndex, CrewAI, DSPy, Haystack, smolagents, Agno, AutoGen, Guardrails) plus MCP. Returns the ones it wired up. Safe to call with none installed.

bryel.shutdown()

Flush pending spans and shut down. Call on process exit.

Feedback and grouping

Group turns into a conversation and attach 👍/👎 by stamping metadata on your spans — set session.id, user.id, and a bryel.interaction.id you mint. See the docs.

Examples

• examples/openai_chat.py — the minimal one call → traced.
• examples/weather_agent.py — a tool-calling agent loop: LLM → tool executions → answer, grouped under one session.

MIT © bryel