orbitrage 0.5.0

One-line observability + intelligent LLM routing for Python agents.

orbitrage

One-line observability + intelligent LLM routing for Python agents.

import orbitrage
orbitrage.init("orb_xxx_yyy")          # one line. that's it.

# now any OpenAI / Anthropic / LangChain call you make is auto-traced
# AND can be auto-routed via:
from openai import OpenAI
client = OpenAI(base_url="https://orbitrage.xyz/api/v1", api_key="orb_xxx_yyy")

What it does

• Zero-latency observability: non-blocking batch span export — your hot path stays at LLM-call speed, not LLM-call + telemetry RTT.
• Auto-instruments OpenAI, Anthropic, LangChain, LlamaIndex, etc. via OpenLLMetry under the hood.
• One endpoint: https://orbitrage.xyz/api/telemetry — auth via your Orbitrage API key.
• Decorators for grouping multi-step workflows:

from orbitrage import workflow, task

@workflow("checkout_flow")
def checkout(user_id):
    plan = planner.invoke(...)
    out  = executor.invoke(...)
    return formatter.invoke(...)

Install

pip install orbitrage

Per-end-user graphs

If your service handles requests for many end-users (your customers' users), pass user_id so the dashboard can partition spans by user inside one workflow:

orbitrage.init("orb_xxx", user_id=current_user.id)

Every span emitted for the rest of the process is tagged with gen_ai.end_user.id. The Orbitrage dashboard uses it to build per-user flow graphs, cost breakdowns, and routing insights.

For long-running servers that switch end-users between requests, call orbitrage.set_user(user_id) at the start of each request.

The workflow your spans show up under is determined by the API key — each key is minted for one workflow from the dashboard. No app_name to pick.

Redacting prompts (opt-out)

Prompt + completion text is captured by default so it renders in the Orbitrage dashboard. To redact content (token counts, costs, and routing decisions still flow):

orbitrage.init("orb_xxx", capture_content=False)