pi.dev is a highly extensible, developer-focused AI agent platform designed for coding and technical workflows. It supports over 15 providers and hundreds of models, including Anthropic, OpenAI, Google, Azure, Amazon Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi, MiniMax, OpenRouter, Ollama, and more. Users can authenticate via API keys or OAuth and seamlessly switch between models during a session using the /model command or Ctrl+L. Favorites can be cycled through with Ctrl+P, and custom providers and models can be added via models.json or third-party extensions. Sessions in pi are organized as tree-structured, shareable histories. Each session is stored as a single file with multiple branches, allowing users to navigate back to any prior point using the /tree command and continue from there. Message types can be filtered, and entries can be labeled as bookmarks. Users can export sessions to HTML using /export or share them instantly as rendered GitHub gists via /share, generating a public URL for easy collaboration. The platform’s strength lies in its extensibility. Rather than embedding features into the core, pi provides primitives—low-level building blocks—so users can construct their own workflows. Extensions are TypeScript modules that have full access to tools, commands, keyboard shortcuts, events, and the entire terminal UI. This enables advanced capabilities like sub-agents, plan mode, permission gates, path protection, SSH execution, sandboxing, MCP integration, custom editors, status bars, and overlays. Yes, even Doom runs on pi. For those who prefer not to build from scratch, pi can generate extensions on request or users can install pre-built packages that match their preferred workflow. Over 50 example extensions are available to inspire and accelerate development. pi’s philosophy centers on minimalism and user control. It avoids baking in features that can be implemented externally—such as MCP support, plan mode, to-do lists, or background bash processes—because these can be added via extensions, scripts, or tools like tmux. This approach keeps the core lightweight while enabling deep customization. Permission prompts are not built in; instead, users can implement their own security flows using extensions that align with their environment. Similarly, plans can be written to files or implemented through custom extensions. The platform prioritizes full observability and direct interaction, allowing developers to see and control every step of the process. For deeper insight into the design decisions, the full rationale is detailed in a dedicated blog post.