Loomkit

A tiny, swappable LLM orchestration core. The seam under an agent harness — without LangChain-scale bloat.

v0 ships the smallest useful surface:

• A Provider Protocol — anything with the right generate() signature is a provider, no inheritance.
• Message / Response / Usage dataclasses.
• One concrete provider: GeminiProvider (official google-genai SDK).

That's it. Streaming, tool calling, agents, retries, and graphs are deliberately deferred — they layer on through the Provider seam without breaking call sites.

Install

pip install loomkit                # core only, zero deps
pip install 'loomkit[gemini]'      # adds google-genai
pip install 'loomkit[gemini,dev]'  # plus pytest

Requires Python 3.10+.

Quickstart

import os
from loomkit import Message
from loomkit.providers.gemini import GeminiProvider

# Reads GEMINI_API_KEY (or GOOGLE_API_KEY) from env; or pass api_key=...
provider = GeminiProvider()

resp = provider.generate(
    [
        Message(role="system", content="You are concise."),
        Message(role="user", content="Name three primary colors."),
    ],
    model="gemini-2.5-flash",
    temperature=0.2,
)

print(resp.text)
print(resp.finish_reason, resp.usage)

Why a Protocol, not a base class

Provider is a typing.Protocol. Adding a new backend means writing one class with a generate() method matching the signature — no imports from loomkit, no inheritance, no registration. Call sites typed against Provider accept any conforming class.

class MyProvider:
    def generate(self, messages, *, model, temperature=None,
                 max_tokens=None, stop=None) -> Response: ...

What's deferred (and how it lands)

Development

pip install -e '.[gemini,dev]'
pytest -q

The Gemini smoke test is skipped unless GEMINI_API_KEY is set.