llm-registry 0.2.0

A centralized registry for discovering and managing LLM model capabilities. Track model features, costs, and limitations across providers like OpenAI and Anthropic. Supports both verified model definitions and user-managed entries with local storage.

LLM Registry

LLM Registry is a Python package that provides a unified interface for discovering and managing the capabilities of various Large Language Models (LLMs). It includes a robust API, a rich CLI, and supports synchronization between local and remote model registries.

Table of Contents

• Overview
• Features
• Installation
• Library Usage
• CLI Usage
• Model Capabilities
• Configuration
• Development
• License

Overview

Manage and discover LLM model capabilities across multiple providers like OpenAI, Anthropic, and more in a centralized registry. Use this package to check model capabilities before initializing provider clients and to manage model metadata efficiently.

Features

• Unified API for capability discovery and management
• Multiple Providers: Supports OpenAI, Anthropic, and others
• Local and Remote Storage: Synchronize your model registries
• Rich Command-Line Interface (CLI)
• Dynamic Capability Management: Easily add, update, and delete model data

Installation

Install via pip:

pip install llm-registry

Library Usage

Integrate the package in your Python projects by following these steps:

Listing Models

from llm_registry import CapabilityRegistry, Provider
registry = CapabilityRegistry()
models = registry.get_models()
for model in models:
    print(model)

Retrieve a Specific Model's Capabilities

model = registry.get_model(Provider.OPENAI, "gpt-4")
if model and model.supports_streaming:
    from openai import OpenAI  # Replace with actual OpenAI client import
    client = OpenAI()  # Initialize client with streaming enabled
    response = client.chat.completions.create(
        model=model.model_id,
        messages=[{"role": "user", "content": "Hello"}],
        stream=True
    )

Add a New Model Capability

from llm_registry.utils import create_model_capability
new_model = create_model_capability(
    model_id="gpt-4",
    provider=Provider.OPENAI,
    model_family="GPT-4",
    input_cost=0.01,
    output_cost=0.03,
    context_window=8192,
    training_cutoff="2023-04",
    supports_streaming=True,
    supports_tools=True,
    supports_json_mode=True,
    supports_system_prompt=True
)

from llm_registry import CapabilityRepository
repo = CapabilityRepository()
repo.save_model_capabilities(new_model)

CLI Usage

The CLI tool llmr allows you to interact with model capabilities directly from the terminal.

List Models

View all available models:

llmr list

*The above screenshot demonstrates how the CLI tool (llmr) currently looks like when listing models.

To filter models by provider:

llmr list --provider openai

Add Model

Add a new model:

llmr add gpt-4 \
    --provider openai \
    --model-family GPT-4 \
    --input-cost 0.01 \
    --output-cost 0.03 \
    --context-window 8192 \
    --training-cutoff 2023-04 \
    --streaming \
    --tools \
    --json-mode \
    --system-prompt

Delete Model

Remove an existing model:

llmr delete gpt-4 --provider openai

Use -f or --force to bypass confirmation.

Model Capabilities

Each model entry tracks:

• Basic Information
  • Provider (e.g., OpenAI, Anthropic)
  • Model ID and Model Family
• Cost Details
  • Input/Output token costs (per 1M tokens)
  • Context window size
  • Training data cutoff date
• Feature Support
  • Streaming responses
  • Tools/Function calling
  • Vision/Image input
  • JSON mode
  • System prompt support

Configuration

Default model data is stored in ~/.llm-registry. You can override the directory by:

• Passing a data_dir parameter to CapabilityRepository in code
• Using the --data-dir option in CLI commands

Development

Requirements

• Python 3.13+
• uv for dependency management

Setup

# Create virtual environment and sync dependencies
uv venv
uv sync --group dev

# Run tests with coverage analysis
pytest -v --cov=llm_registry

License

Distributed under the MIT License. See LICENSE for more information.