Pi-tree cuts token usage by 80% with tree-structured conversations

Pi-tree is a self-hosted AI reading companion that replaces flat chat threads with navigable trees [Dev.to]. The project ships as an AGPL-3.0 Docker image (ghcr.io/shuowu/pi-tree:latest) and includes a TypeScript SDK for custom extensions [GitHub]. The core server runs on any OpenAI-compatible endpoint, such as OpenAI, Gemini, Claude, DeepSeek, or local Ollama, allowing a 12B parameter model to handle most reading tasks.

Pi-tree's tree-structured conversations reduce token usage by isolating each branch, resulting in an 80% decrease in token volume and associated API cost [Dev.to]. This approach also lowers hallucination rates by limiting each branch's context to its path from the root. Additionally, Pi-tree equips the model with deterministic tools (get_latest_rss, search_papers, get_youtube_transcript) that fetch exact sections on demand, making it an on-call researcher rather than a probabilistic retriever.

The tool includes built-in parsers for EPUB, PDF, MOBI, arXiv papers, RSS feeds, and YouTube transcripts. Developers can add new source types via the plugin SDK (definePiTreeExtension) and expose them as tools the model can call on demand [GitHub].