mindroot 0.4.0

An agent framework using LLMs

Mindroot

Note: some of the following is a work in progress and not yet functional. A few of them aren't really started.

Mindroot is a powerful plugin-based Python framework for creating, deploying, and sharing AI agents and related models. It offers a flexible architecture with indices and a public registry (coming soon) for easily sharing and finding plugins, agents, personas, services, knowledgebases, and apps.

Installation

You can install Mindroot using pip:

(You probably want to create a virtual environment first: python -m venv venv and source venv/bin/activate)

pip install mindroot

For development, you can install the package in editable mode:

git clone https://github.com/mindroot/mindroot.git
cd mindroot
pip install -e .

Configuration

You will need to set environment variables like JWT_SECRET_KEY (anything you want) and LLM provider keys etc.

To enable email verification: env. variable REQUIRE_EMAIL_VERIFY=true

See SMTP core plugin readme

Starting the Server

Go to the directory with the virtual environment and run:

mindroot

or if you want to start the server on a different port:

mindroot -p 8001

Note that you will always need to start from that working directory if you want your settings to be preserved.

Installing Plugins

Most of the commands or services are in plugins that you have to install. Go to the /admin page and find the Plugins section Click on the 'default' Index It should show a list of Plugins I suggest installing all of them.

The first would be either Anthropic (recommended) or OpenAI

Important: You will need either ANTHROPIC_API_KEY or OPENAI_API_KEY set in your environment variables.

To install a plugin just click the Install button and wait 10-20 seconds and you will see a confirmation pop up.

Important: You will need to restart the server to see the changes.

For now I recommend not using the Server control section on the Admin page but rather just hit Control-C in the terminal and then run mindroot again.

Developers: Creating Plugins and Tool Commands etc.

See the Plugin Documentation

Adding Commands to Agents

Go to Admin -> Agents Select an agent from the drop down, such as Assistant

There are toggle switches for all of the commands. I recommend avoiding the 'replace_' commands in files for now because they are not reliable.

Key Features:

• Public registry for sharing and finding plugins, agents, personas, models, and knowledgebases
  • https://registry.agenthost.org (work in progress)
  • Customizable and swappable for user-specific registries
• Extensible plugin architecture for adding services, commands, and building arbitrary web apps
• Customizable AI agents with persona definitions
• Intelligent service management based on agent requirements
• Flexible service providers for various AI capabilities
• Plugins can add/use hooks and pipelines such as for modifying prompts, running startup commands, or anything you want
• Easily customizable UI built on Jinja2 and Lit Web Components
• Support for both local and remote AI services
• RAG: easily share, find and use pre-generated embeddings and documents for topic knowledgebases

Core Concepts Overview

mindroot revolves around several key concepts:

1. Open Public Registry: A flexible system for indexing and sharing plugins, agents, personas, and potentially models. It can be customized or replaced with user-specific registries.

2. Plugins: Extend the functionality of mindroot, providing new features, services, or integrations.

3. Agents and Personas: AI agents with defined capabilities and customizable personalities.

4. Services and Providers: Backend services that power agent capabilities, with support for swapping between local and remote implementations.

5. Intelligent Service Management: The system automatically determines and installs required services based on agent definitions.

6. UI Customization: Easily modifiable user interface through theme overrides and injections.

7. RAG and Knowledgebases: The community can easily share and search for topic embeddings and document sets rather than everyone rebuilding them per topic.