Pi AI Agent Toolkit

Build powerful AI agents with a unified LLM API, agent runtime, and interactive terminal UI

Welcome to Pi

Pi is a comprehensive toolkit for building AI agents with full control over your workflows. From a flagship interactive coding agent CLI to low-level APIs for custom implementations, Pi gives you the building blocks to create intelligent, tool-using applications.

Key Philosophy: Aggressively extensible. Features that other tools bake in can be built with extensions, skills, or installed from third-party packages. This keeps the core minimal while letting you shape Pi to fit how you work.

What’s Included

Pi is a monorepo containing seven packages, each solving a specific problem in the AI agent stack:

Coding Agent

Interactive CLI tool with read, write, edit, and bash tools. Supports extensions, skills, themes, and packages.

LLM API (@mariozechner/pi-ai)

Unified API for 15+ LLM providers with automatic model discovery, tool calling, and cross-provider handoffs.

Agent Runtime (@mariozechner/pi-agent-core)

Stateful agent with tool execution, event streaming, and transport abstraction for building custom agents.

Terminal UI (@mariozechner/pi-tui)

Differential rendering framework for flicker-free terminal interfaces with built-in components.

Web UI (@mariozechner/pi-web-ui)

Web components for AI chat interfaces with attachments, artifacts, and storage.

Slack Bot (@mariozechner/pi-mom)

Slack bot that delegates messages to the pi coding agent for team collaboration.

vLLM Management (@mariozechner/pi-pods)

CLI for managing vLLM deployments on GPU pods for self-hosted inference.

Core Features

Unified LLM Access

Work with any LLM provider through a single, consistent API:

• 15+ Providers: OpenAI, Anthropic, Google, Mistral, xAI, Groq, Cerebras, OpenRouter, and more
• Automatic Discovery: Models and capabilities detected automatically
• Tool Calling: TypeBox-validated function calling across all providers
• Cross-Provider Handoffs: Switch models mid-conversation while preserving context
• OAuth Support: Built-in support for subscription-based services (Claude Pro, ChatGPT Plus, GitHub Copilot)

Agent Building Blocks

Build custom agents with minimal boilerplate:

• Event-Driven: Subscribe to agent lifecycle events for real-time UI updates
• Tool Execution: Automatic validation and error handling for tool calls
• State Management: Built-in message history and context management
• Steering & Follow-up: Interrupt agents mid-execution or queue follow-up tasks
• Transport Abstraction: Run agents over HTTP, WebSocket, or in-process

Terminal & Web UIs

Pre-built components for both terminal and browser:

• Terminal: Editor, markdown renderer, select lists, loaders, overlays
• Web: Chat panels, artifact viewers, file attachments, session management
• Themes: Customizable styling for both environments
• Differential Rendering: Efficient updates with no flicker

Use Cases

Interactive Coding Assistant

The flagship @mariozechner/pi-coding-agent provides a full-featured terminal interface for AI-assisted development with file operations, bash execution, and extensibility through TypeScript modules.

Custom Agent Applications

Use @mariozechner/pi-agent-core and @mariozechner/pi-ai to build domain-specific agents with custom tools, prompts, and workflows.

Web-Based Chat Interfaces

Deploy @mariozechner/pi-web-ui components in React/Next.js apps for AI chat with document processing and artifact generation.

Team Collaboration via Slack

Run @mariozechner/pi-mom to bring AI assistance into Slack channels for team-wide access.

Self-Hosted LLM Infrastructure

Manage GPU pods and vLLM deployments with @mariozechner/pi-pods for cost-effective inference.

Quick Navigation

Installation

Install packages and set up your environment

Quick Start

Get running in under 5 minutes

Architecture

Understand how the pieces fit together

LLM Providers

Connect to 15+ LLM providers

Building Extensions

Extend the coding agent with TypeScript

API Reference

Explore the complete API surface

Community & Support

Discord Community

Join the community for support and discussion

GitHub Repository

View source code, report issues, and contribute

Design Principles

Minimal Core, Maximum Extensibility: Pi deliberately omits features like sub-agents, plan mode, and permission popups from the core. Instead, these can be built as extensions, installed from packages, or implemented however fits your workflow. This keeps the foundation small and fast while enabling infinite customization.

No Lock-in: Context and messages are plain JSON. Switch between models mid-conversation. Use different providers for different tasks. Pi doesn’t force you into a single provider or workflow.

What’s Next?

1. Install Pi

   Follow the installation guide to set up the packages you need

2. Run the Quick Start

   Get hands-on with the quick start tutorial and build your first agent interaction

3. Explore the Docs

   Dive into core concepts or jump straight to API reference