toki 0.2.6rc1

Minimal universal LLM interface API

Toki

A minimal python library for talking to LLMs.

When using this library, in most cases one can simply switch the model name (e.g. 'openai/gpt-5' -> 'google/gemini-2.5-pro') and everything will just work.

Getting Started

pip install toki

from toki import Model, Agent, get_openrouter_api_key
from easyrepl import REPL

agent = Agent(
    Model('openai/gpt-5', get_openrouter_api_key())
)

for query in REPL():
    agent.add_user_message(query)
    for chunk in agent.execute(stream=True):
        print(chunk, end='', flush=True)
    print()

Features

• Model: stateless class for talking to LLMs. includes blocking and streaming
• Agent: class for maintaining conversation history when talking with a model
• StateMachine/ClassStateMachine: tools for easily implementing complex agentic interaction scenarios

Backends

The primary backend is OpenRouter which provides REST API access to all major LLM models on the market

Issues

OpenAI models fail to work with tools

• TODO: consider adding a separate backend just for openai. also consider adding a text only tools interface to skip around the issue

Dev

Rebuild models list file

toki-fetch-models

• TODO: look into making this step part of the github action for publishing new version (or periodically checking the models list and auto updating/building new versions)