Pi Coding Agent

• Running a coding task in a repository, then branching back to an earlier message to try a different approach without losing the first attempt
• Switching between a frontier model and a cheaper or local model within one session
• Scripting Pi in print or JSON mode as a step in CI or a batch job
• Embedding the agent in an internal tool through the TypeScript SDK or RPC mode
• Building a team package that adds custom tools, skills, and prompt templates on top of the core agent

Pi is for developers who find full-featured coding agents too opinionated. The core is small, model-agnostic, and easy to script, and its tree-structured sessions let you try an alternative approach without losing the first one. The cost is that you assemble the rest yourself. There are no approval prompts, no built-in MCP or sub-agents, and no IDE integration, so it rewards people willing to write extensions and run it in a container. Teams that want guardrails and batteries included should look elsewhere.

About Pi Coding Agent

Pi takes the opposite approach to batteries-included coding agents. The core agent inspects files, runs commands, and edits code, and it shows every read, search, command, and edit in the transcript. It does not ask for approval before each tool call. Pi deliberately leaves out MCP, sub-agents, permission popups, plan mode, built-in to-dos, and background bash. You add those through TypeScript extensions, or use plain tools such as tmux and TODO.md files. Sessions are stored as a tree, so you can return to an earlier message and branch with /tree, split off separate work with /fork, or copy the current state with /clone. When you leave a branch, Pi can summarize it into the one you enter. You can switch models mid-session, steer the agent while it works, and shape context with AGENTS.md, SYSTEM.md, and compaction. Pi works with a long list of providers through API keys, including Anthropic, OpenAI, Google Gemini, DeepSeek, Mistral, Groq, xAI, OpenRouter, Kimi, and MiniMax. It also supports cloud platforms such as Azure OpenAI, Amazon Bedrock, and Google Vertex AI, OAuth sign-in for supported subscriptions such as OpenAI Codex and GitHub Copilot, and local models through llama.cpp or OpenAI-compatible servers such as Ollama and LM Studio. Beyond the interactive TUI, Pi runs in print, JSON event stream, and RPC modes and as a TypeScript SDK. The building blocks are published as separate npm packages (pi-ai for a unified multi-provider LLM API, pi-agent-core for the agent runtime, and pi-tui for the terminal UI), and other agents use them: DeepSeek Harness routes third-party providers through an adapter built on pi-ai, and Kimi Code's TUI is built on pi-tui. The cost of this minimalism is safety and convenience. Pi runs with your user's permissions and has no built-in permission system, so the project recommends a container, micro-VM, or sandbox. Features other agents ship by default have to be added as packages.