pi-herdsman

Pi is a terminal‑based coding agent that runs as a minimal use you can extend with your own tools, prompts and UI tweaks. It focuses on letting you adapt the agent to your workflow rather than forcing a preset set of features. The project is distributed via a shell installer or npm‑style packages such as @earendil-works/pi-coding-agent.

You interact with Pi through a text user interface (TUI) that supports four operating modes: an interactive full‑screen view, a print/JSON mode for generating scripts, an RPC protocol for non‑Node integrations, and an SDK for embedding the agent in other applications. Commands are entered with Enter to steer the current run, or Alt‑Enter to queue a follow‑up after the current tool finishes. The interface also lets you branch a session tree, rewind to any prior message, and export or share the whole session as HTML, a GitHub gist, or a shareable URL.

Pi's extensibility is built around extensions, skills and prompt templates that you can load on demand. Extensions are TypeScript modules that can add custom commands, keyboard shortcuts, tool integrations or UI elements such as overlays. Skills are capability packages that bring instructions and tools into the context without inflating the prompt cache. Prompt templates are reusable Markdown files you invoke with a slash command. The system also supports loading project‑specific instructions from AGENTS.md or SYSTEM.md files placed in your home directory or project folders.

The agent can work with more than fifteen model providers, including Anthropic, OpenAI, Google, Azure, Bedrock and several open‑source options. Authentication is handled via API keys or OAuth, and you can switch models mid‑session with /model or Ctrl + L. A built‑in compaction feature automatically summarizes older messages when the context window approaches its limit, and you can replace that logic with custom summarizers if needed. Sessions are stored as a single tree‑structured file, making it easy to filter, bookmark, or branch history.

Key features

• Four interaction modes: interactive TUI, print/JSON, RPC, SDK
• Extensible via TypeScript extensions, skills and prompt templates
• Session tree storage with branching, bookmarking and export options
• Support for 15+ model providers and mid‑session model switching
• Automatic context compaction with optional custom summarizers
• Command shortcuts for steering (Enter) and queuing follow‑ups (Alt‑Enter)