Pi Agent

Pi Agent

Pi Agent is a terminal-based coding agent built by Earendil Inc. and hosted at https://pi.dev/, with the philosophy that the harness should adapt to the developer. Pi Agent is a TUI for the terminal, supports tree-structured sessions, ships skills and project context files, and uses a minimal default system prompt for token efficiency. The tagline is "There are many agent harnesses, but this one is yours."

The product ships in four modes. Interactive mode delivers the full TUI with rich keyboard shortcuts and themable visuals. Print mode runs Pi Agent headlessly from a shell script with a single query. JSON mode streams structured events from a run for piping into other tools. RPC mode exposes Pi Agent over a JSON protocol on stdin and stdout for non-Node integrations. An SDK is provided for embedding Pi Agent inside other applications, with OpenClaw referenced as a real-world SDK user.

Pi Agent is model-agnostic. It supports Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi For Coding, MiniMax, NVIDIA, OpenRouter and Ollama, with more providers added regularly. Authentication works via API keys or OAuth. /model or Ctrl+L switches models mid-session, Ctrl+P cycles through favourites, and new providers can be registered through a models.json file or via extensions.

Context engineering is a first-class concern. Pi Agent loads AGENTS.md project instructions from ~/.pi/agent, parent directories and the working directory. SYSTEM.md can override or extend the default system prompt on a per-project basis. Compaction auto-summarises older messages as the context window fills. Skills are capability packages loaded on demand. Prompt templates are reusable Markdown prompts invoked with /name. Extensions, written in TypeScript, can inject context before each turn, filter history, implement RAG or build long-term memory. Sessions live as trees; /tree navigates, /export writes HTML, /share pushes to a GitHub gist for a shareable URL.

Pi Agent is aimed at developers who want their coding agent to bend to their workflow instead of bending their workflow around the agent.