PI

Overview:

Pi is a minimal terminal coding harness for interacting with large language models (LLMs). It focuses on providing a core set of powerful defaults while allowing developers to extend and customize every aspect of its behavior through TypeScript Extensions, Skills, and Prompt Templates. Unlike some other coding agents, Pi intentionally omits built-in features like sub-agents, plan mode, or permission popups, instead providing the mechanisms to build or install these capabilities as needed. It is designed for developers who want a flexible, scriptable, and extensible interface for AI-assisted coding.

Core Features:

• TypeScript Extensions: Extend Pi with custom tools, commands, keyboard shortcuts, event handlers, and UI components. Extensions can replace built-in tools, add sub-agents, implement custom compaction, and more.

• Skills: On-demand capability packages that follow the Agent Skills standard. Invoked via /skill:name or loaded automatically, allowing the agent to perform specialized tasks.

• Prompt Templates: Reusable Markdown files that can be expanded via /templatename, enabling developers to quickly inject pre-written instructions or workflows.

• Session Branching: Sessions are stored as JSONL files with a tree structure, allowing in-place navigation, forking, and cloning. Use /tree to navigate history, /fork to create new sessions from a previous point, and /clone to duplicate the active branch.

• Multiple Execution Modes: Runs in interactive, print (-p), JSON (--mode json), RPC (--mode rpc), and SDK modes (createAgentSessionRuntime()), enabling integration into various processes and applications.

• Context File System: Loads AGENTS.md or CLAUDE.md files from global and project directories to provide project-specific instructions and conventions.

Use Cases:

• Developing a coding agent that matches a personal workflow: A developer can use Pi to build a custom coding agent by writing extensions for specific tool usage (e.g., git auto-commit, SSH execution) and disable built-in tools they don't need.

• Integrating an LLM agent into a larger application: A developer can use Pi's SDK (createAgentSessionRuntime()) or RPC mode to embed an agent within a custom IDE, CI/CD pipeline, or other tooling.

• Sharing agent configurations across a team: A team can create and share Pi Packages (via npm or git) containing their custom extensions, skills, and prompt templates to standardize their coding agent workflow.

• Experimenting with different LLM providers and models: Developers can easily switch between multiple providers (Anthropic, OpenAI, Google, etc.) and models via a command or keyboard shortcut, allowing rapid experimentation.

Why It Matters:

Pi stands out as an open-source coding harness that prioritizes extensibility over a fixed feature set. By providing a minimal core and a powerful extension mechanism, it avoids locking users into a specific workflow. Developers can use TypeScript to add or replace any functionality, from sub-agents and plan modes to custom UI components and MCP server integration. This approach allows for a high degree of customization and integration for teams that want to build bespoke AI-assisted development environments without forking the core tool.