Python SDK

The SDK is open source: github.com/llmjury/llmjury-python (Apache-2.0 — issues and PRs welcome).

Install from PyPI (Python 3.8+, zero runtime dependencies — the transport is stdlib urllib; the package installs as llmjury-sdk and imports as llmjury):

pip install llmjury-sdk

Set up your API key

The client won't start without your org's publishable key (llmj_pk_…) — grab it from the dashboard under Settings → API keys (admins can view it there any time). Then either:

Recommended — set the environment variable. The client reads LLMJURY_API_KEY automatically, so the key never lives in your code:

export LLMJURY_API_KEY=llmj_pk_...   # shell / CI secret / deployment env / .env file

from llmjury import Client

client = Client()   # picks up LLMJURY_API_KEY — nothing else to configure

Or pass it explicitly if your platform injects secrets another way:

client = Client(api_key="llmj_pk_...")

If neither is provided, Client() raises immediately with a message saying exactly that — you can't accidentally run unauthenticated. The key is sent as X-API-Key on every request. The SDK defaults to the production API; point it at a local stack with LLMJURY_BASE_URL (or base_url=).

Quickstart

from llmjury import Client

# ---- setup, once at startup -------------------------------------------------
# Client() needs your publishable API key (dashboard -> Settings -> API keys).
# Easiest: `export LLMJURY_API_KEY=llmj_pk_...` — it's read automatically.
# Or pass it in code: Client(api_key="llmj_pk_...", ...)
client = Client(experiments=["checkout-prompt"])          # prefetch by experiment NAME
llm = client.wrap(anthropic_client, "checkout-prompt")    # every model call is now traced

# ---- per request ------------------------------------------------------------
with client.as_user(user_id):
    # Variant prompt from client memory; your in-code default survives an outage.
    p = client.get_prompt("checkout-prompt", user_id, default="You are a helpful assistant.")
    # Call your provider client DIRECTLY — latency/tokens/errors are intercepted.
    response = llm.messages.create(model="claude-haiku-4-5", max_tokens=1024,
                                   system=p.prompt,
                                   messages=[{"role": "user", "content": user_input}])

# The ONLY explicit metric: your business outcome.
client.track("business_event", {"experiment_id": "checkout-prompt", "user_id": user_id,
                                "variant": p.variant, "business_metric": "conversion", "value": 1})

What the client does

• get_prompt(experiment, user, default) — the recommended entrypoint: resolves the user straight to their variant's prompt text. default is your in-code fallback, returned whenever the config isn't cached, assignment fails, or the variant has no prompt — the app always has a working prompt, even during a full LLMJury outage. Returns (variant, prompt, fallback).
• get_variables(experiment, user, defaults) — same idea for custom variables (model, temperature, …): the variant's configured values merged over your in-code defaults, from client memory — never an API call on the request path.
• wrap(provider_client, experiment) + as_user(user) — setup-once interception: wrap your OpenAI/Anthropic-style client at startup, bind the user per request with with client.as_user(user_id):, then call the provider directly. Every model call records the exposure plus a model_call event with measured latency, tokens, model, and errors — no call-site code. (intercept_model_call is the explicit context-manager form.)
• assign(experiment, user) — the low-level call: deterministic local bucketing against the polled config; returns the variant key, or None until the config loads — treat None as "use control". See bucketing.
• track(event, payload) — enqueue-only; a background thread flushes batches with bounded retries, then spills to an optional offline file (replayed within 24h with original timestamps) or drops with a log line. It never blocks or raises into your call path. With wrap(), the only event you track by hand is your business_event.
• Experiments are addressed by unique name or id — names resolve through the polled config and bucketing always runs on the canonical id, so both address forms assign identically.
• Config is polled from GET /v1/config (ETag/304, 60s) and refreshed immediately when an ingest ack reports a newer config_version. Async variants: aassign / atrack.

Full details in the SDK README.