Pi is an open-source, minimalist coding agent built by Mario Zechner that runs in the terminal with an intentionally lean design, shipping with only four core tools (read, write, edit, and bash) and a system prompt of roughly 300 words. It solves the problem of bloated AI coding tools by taking a radically minimal approach where the agent itself is kept simple and lightweight, while extensibility is achieved through TypeScript extensions, skills, and prompt templates that developers build to match their specific needs. Pi treats the context window as a scarce resource, keeping system overhead minimal to maximize the space available for actual code and project information.
Pi differentiates itself through its extreme minimalism and extensibility philosophy, deliberately omitting features like sub-agents, plan mode, and MCP support that other tools bundle by default, instead encouraging users to build exactly the capabilities they need. The agent runs in four modes: interactive chat, print/JSON output, RPC for process integration, and an SDK for embedding in custom applications. Sessions are stored as navigable trees, extensions can be packaged and shared via npm or Git, and the platform supports all major LLM providers including Anthropic, OpenAI, Google, Azure, Mistral, and Groq.
Pi targets developers who prefer tools that do one thing well and can be composed with other Unix-style utilities, rather than monolithic platforms that try to handle every development task. It is especially appealing to experienced developers who want full control over their AI agent's behavior and are comfortable writing TypeScript extensions to add domain-specific capabilities. Pi's lightweight architecture makes it practical for resource-constrained environments and developers who want to experiment with different agent designs without the overhead of larger frameworks.