pi is an open-source AI agent toolkit that bundles a lot into one repo: a coding agent CLI, a unified multi-provider LLM API (OpenAI, Anthropic, Google), TUI and web UI libraries, a Slack bot, and vLLM pods for self-hosting.
## The unified-API core
At the center is a single API that abstracts across providers — switch models without rewriting integration code. Authenticate with `/login` for subscription providers, or set an API key like ANTHROPIC_API_KEY. The coding agent supports interactive mode, slash commands, context files, session management, branching, tree navigation, and context compaction.
## Designed to be extended
The pitch is aggressive extensibility — pi doesn’t dictate your workflow. It exposes TypeScript modules for tools, commands, events, and custom UI, so you build your own agent on top rather than living inside someone else’s product. The vLLM pods piece means you can run the whole stack on your own inference infrastructure.
## Why it matters
Most agent tooling is either a finished product (Claude Code, Cursor) or a low-level SDK. pi sits in between — a toolkit with batteries included (CLI, UI libs, Slack bot) that you can still take apart and rebuild. For teams that want to ship a custom internal agent without starting from zero, that middle ground is exactly the gap.

Leave a comment