bitfrost

Drop-in OpenTelemetry observability for Python LLM apps. Standalone or with any OTLP backend — runs on your machine, sends nothing unless you ask it to.

Bitfrost turns the OpenTelemetry spans your LLM libraries already emit into a clean, queryable event stream — in your terminal, in a local web dashboard, in SQLite, or shipped to any backend you choose. One line to start, zero accounts required.

import bitfrost

bitfrost.quickstart(agent="my-app")   # auto-detects openai / anthropic / litellm / smolagents
# ...your normal LLM calls now stream to the terminal, color-coded, with tokens + cost.

bitfrost serve capture.db — a local, offline dashboard. No account, no upload.

Why bitfrost

• Standalone first. Capture to your terminal, a JSONL file, or SQLite with zero configuration and zero network calls. Voight is one optional backend, never a requirement.
• Drop-in. Built on the OpenTelemetry GenAI semantic conventions, so it works with the instrumentation libraries you already use — across both the v1.27 and v1.32+ attribute generations.
• Batteries included. Five backends, four auto-instrument helpers, a rich CLI, an interactive TUI, and an offline web dashboard.
• Private by default. Three privacy levels with PII scrubbing (12 patterns + Luhn) applied before any event leaves the process.
• Never crashes your app. Every failure path degrades gracefully — a dead endpoint or a missing optional dependency never takes down your code.

Install

pip install bitfrost                 # core
pip install 'bitfrost[cli,serve]'    # + CLI, TUI and local web dashboard
pip install 'bitfrost[all]'          # everything

Python 3.10–3.13.

Capture anywhere

Pick a backend and pass it to any instrument helper (or quickstart). All concrete backends live under bitfrost.backends.*:

import bitfrost
from bitfrost.backends.sqlite import SQLiteBackend

bitfrost.instrument_openai(backend=SQLiteBackend("events.db"), agent="my-app")
# then:  bitfrost serve events.db

Fan out to several at once:

import bitfrost
from bitfrost.backends.tee import TeeBackend
from bitfrost.backends.console import ConsoleBackend
from bitfrost.backends.sqlite import SQLiteBackend

bitfrost.instrument_auto(
    backend=TeeBackend(ConsoleBackend(), SQLiteBackend("events.db")),
    agent="my-app",
)

Auto-instrument helpers

import bitfrost

bitfrost.instrument_openai()       # opentelemetry-instrumentation-openai
bitfrost.instrument_anthropic()    # opentelemetry-instrumentation-anthropic
bitfrost.instrument_litellm()      # litellm CustomLogger adapter
bitfrost.instrument_smolagents()   # openinference-instrumentation-smolagents
bitfrost.instrument_auto()         # instrument every supported lib that's installed

Already wiring OpenTelemetry yourself? Skip the helpers and attach the exporter to your own TracerProvider:

from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from bitfrost.exporter import BitfrostExporter
from bitfrost.backends.console import ConsoleBackend

provider = TracerProvider()
provider.add_span_processor(
    BatchSpanProcessor(BitfrostExporter(ConsoleBackend(), agent="my-app"))
)

CLI

bitfrost watch  capture.db        # live tail, one styled line per event
bitfrost replay capture.jsonl     # re-render a captured run start to finish
bitfrost query  capture.db "SELECT model, COUNT(*) FROM events GROUP BY model"
bitfrost vacuum capture.db --keep-days 7
bitfrost tui    capture.db        # full-screen interactive dashboard
bitfrost serve  capture.db        # local web dashboard at http://127.0.0.1:8080

bitfrost tui — the same capture, navigable in your terminal. Prompts masked by default.

Privacy

Three levels, applied in-process before any event is sent:

• minimal — metadata only, no prompt/response content
• standard (default) — content kept, PII scrubbed (12 patterns + credit-card Luhn check)
• full — everything verbatim (use only for local debugging)

bitfrost.instrument_openai(privacy="minimal")

Standalone, or with Voight

Bitfrost is useful entirely on its own — the core has no dependency on any hosted service. If you want managed dashboards, team sharing, and per-user cost attribution, the optional VoightBackend ships your events to voight.xyz. Everything else stays exactly the same.

Documentation

• Cookbook — task-oriented recipes
• Write your own backend — the ExportBackend protocol
• Full docs: docs.voight.xyz/python/bitfrost

License

MIT. Built by Voight.