Pi is a minimal, terminal-based agent harness designed to adapt to your specific development workflows rather than imposing a rigid structure. Developed by Earendil Inc., it functions as a highly extensible coding agent that gives users granular control over their environment. Unlike many other AI agent tools, Pi ships with powerful core defaults while deliberately omitting features like built-in sub-agents, plan modes, or rigid permission popups. Instead, these functionalities can be added via a robust system of extensions, skills, and prompt templates, allowing developers to build the workflow that best suits them.

Pi operates by treating its own functionality as primitives. If a developer needs a specific tool, command, UI tweak, or workflow pattern, they can build it as a TypeScript module extension or install a package shared by the community via npm or git. This design philosophy ensures the core agent remains lean and performant, avoiding the bloat common in other coding agents. The agent supports over 15 AI providers, including OpenAI, Anthropic, Google, and many others, allowing seamless switching between models mid-session to optimize performance and cost.

Some of the key features are:

• Extensibility: Build custom workflows, tools, and UI components using a full TypeScript extension system.
• Model Versatility: Switch between hundreds of supported models from multiple providers like Anthropic, OpenAI, Mistral, and Ollama in real-time.
• Session Management: Native support for tree-structured history, allowing branching and navigating through past session states.
• Context Engineering: Customizable system prompts and project instructions per directory, with intelligent context compaction.
• Multiple Operational Modes: Operates in interactive TUI, print/JSON mode, RPC protocol, or as an embedded SDK.
• Progressive Disclosure: Load capabilities like skills and instructions on-demand to maintain prompt cache efficiency.
• Shareable Sessions: Export entire session trees as HTML or shareable URLs for collaboration.

Operationally, Pi provides a terminal-based interface where developers can interact directly with the agent. The agent can be configured at the project level using AGENTS.md and SYSTEM.md files, which define project-specific instructions and system prompts. Users control the agent via slash commands, and can steer the agent's actions in real-time. The agent maintains a persistent session log that tracks every interaction and state change, which can be viewed or resumed at any point.

Some common use cases include:

• Custom Agent Workflows: Creating bespoke development environments that enforce specific coding standards or CI/CD integrations.
• Multi-Model Experiments: Comparing the performance of different LLMs on the same coding tasks by switching them within a single session.
• Project-Specific Automation: Loading project-specific skills and tools that are automatically available when opening a particular directory.
• Observability and Collaboration: Branching session history to explore different solutions to a bug or feature, then sharing the successful path with teammates.