CAELUM

Open source · Apache 2.0 · Python SDK

The control plane for AI inference

Drop-in replacement for the OpenAI, Anthropic, Google Gemini and Mistral Python SDKs. Cache, route, observe and fall back across providers, by changing one import.

Get startedView on GitHub

Get started in 30 seconds

terminal
pip install "caelum-ai[openai]"
Before
from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)
After: CAELUM
from caelum import OpenAI   # <- the only change

client = OpenAI(cache=True, retry=True)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)
client.metrics.print_summary()  # requests, cache hits, cost

Your code, your types and your API keys stay the same. Caching and retries are opt-in. CAELUM never resells tokens: you keep your own provider accounts.

Why CAELUM

Exact & semantic cache

Identical requests are answered locally from SQLite. An opt-in semantic cache reuses answers to rephrased questions, after a small model confirms both ask the same thing.

Routing by config

Send requests to different models from a dict, JSON or YAML file. OpenAI ⇄ Anthropic request translation built in.

Built-in observability

client.metrics tracks calls, tokens, cache hits, latency and estimated cost. Export to JSON.

Multi-provider fallback

When a provider fails, retry with backoff, then fall back to the model you chose on another provider, or to a free local model (Ollama, llama.cpp, vLLM, LM Studio).

30%
of calls saved on a benchmark with 30% repeated requests
< 1 ms
p95 overhead measured for the exact cache (target: under 5 ms)
0 / 50
look-alike questions answered wrongly by the semantic cache, with 37 of 50 rephrasings reused (real API test)
4
SDKs wrapped: OpenAI, Anthropic, Google Gemini, Mistral, plus local models

Figures come from our own benchmarks (see the repository). Your savings depend on how often your prompts repeat.

CAELUM Cloud is coming

A hosted dashboard for your inference costs and cache. Join the waitlist to hear first.

Waitlist opens soon.