Combo Draft: Unified Model Access & Observability

Scenario

The pain of multi-model access isn't the models themselves, it's the "access surface": each service configures its own key and connects to its own vendor directly, costs scattered and invisible. This combo chains "access → observability" into a deterministic pipeline: litellm gateway as the unified entry (one base_url for all models, keys centrally managed), logfire for OpenTelemetry observability (calls/costs/latency at a glance). Each tool owns one stage, no overlap.

Combo members

litellm: unified access gateway (entry side, newly recorded)

Download: PyPI litellm[proxy] (official; China Tsinghua mirror)
Version: >= 1.98.0 (measured 2026-08-29; Python >= 3.8)
Install:

python3 -m venv ~/.litellm-venv
~/.litellm-venv/bin/pip install -i https://pypi.tuna.tsinghua.edu.cn/simple litellm[proxy]

logfire: observability (observation side, existing entry)

Download: PyPI logfire (official; China Tsinghua mirror)
Version: == 4.41.0 (measured version recorded 2026-08-28)
Install:

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple logfire==4.41.0

Install order

litellm (gateway) → 2. logfire (observability); the combo is ready only when all are installed and running. litellm is a server process (default port 4000), logfire is SDK/CLI form.

Dependency warnings

Python >= 3.8 (both tools pip form); litellm recommends an isolated venv (many dependencies)
litellm's first-start remote cost-map pull times out and auto-falls back to the local backup (China network observation, measured 2026-08-29)
Port: litellm defaults to 4000 (--port can change); conflicts with services occupying 4000
Conflicts: the two tools' write surfaces don't overlap (proxy config / telemetry data) → compatible (combo measured state: 2026-08-29 litellm proxy surface smoke passed)
Note: real forwarding needs model API keys (keys go centrally into the litellm config, clients no longer scatter them)

Acceptance check (one shot)

Run this block after installing; all green = combo ready; any failure -> back to the matching step in the install walkthrough.

~/.litellm-venv/bin/litellm --version
logfire --version

预期输出: litellm >= 1.98.0; logfire 4.41.0. Command surface green = ready; live-forwarding acceptance: configure keys, start the proxy, curl http://localhost:4000/health/liveliness returns "I'm alive!" (measured)

Note: live multi-model forwarding needs per-model API keys (centralized in the litellm config); not measured, honestly marked.

Verification evidence

Every combo is measured in an isolated environment (smoke + write-surface check); unmeasured items are honestly marked in the text, and measurement dates are listed in the verification section below.

Verified 2026-08-29 · TTL expires — (180 days)

2026-08-29: litellm 1.98.0 measured (Tsinghua mirror venv install 80s; openai/mock + dummy key started the proxy: /health/liveliness "I'm alive!" + /v1/models mock-gpt registered successfully; server stopped cleanly)
2026-08-28: logfire recorded and measured (install/command surface)
Full chain (gateway → observability end-to-end) to be measured at combo acceptance (litellm proxy surface verified; the joining surface = real key forwarding + logfire telemetry access)

Combo Draft: Unified Model Access & Observability Partial