P-ai：支持多智能体协作的桌面 AI 工作系统

A ready-to-use self-growing desktop AI assistant for long-running tasks, memory, agents, tool reviews, MCP, and high-concurrency workspac…

P-ai (PAI) Languages / 语言 简体中文 | 繁體中文 | English | 日本語 A self-growing desktop AI work system — ready-to-use, with agent delegation, long-term memory, tool review, MCP, and high-concurrency workspace automation. PAI is an actively evolving desktop AI work system. It is not a chat client — it is a complete desktop system organized around conversations, tasks, memory, persona organization, tools, review, and remote messaging. The backend uses Rust async concurrency and streaming architecture to guarantee response speed; the frontend uses Vue 3 + DaisyUI for a clean interface. All data is stored locally, with no intermediate servers. The same chat runtime is also available in the browser and as a VS Code sidebar. Host differences stay in the transport layer; the product itself is not desktop-only. Entry & Efficiency Global hotkey summon, Windows background voice wake (hold the record key), quick screenshot — PAI brings desktop AI access to "summon anytime, handle anything, continue anywhere." Supports local sessions, remote sessions, and multiple parallel sessions; quick commands can trigger common operations in one keystroke. Organization & Personas Multiple personas can be independently configured, each with its own avatar and private memory, and they form an organization through parent-child relationships. Tasks and sessions are separated by persona, identity, and responsibility. Local sessions support multi-agent group chat; remote sessions support WeChat, Feishu, DingTalk, OneBot, QQ Official, and other protocols. Interface & Interaction UI, chat style, colors, and fonts are all customizable, with multiple windows running in parallel. Fast response, clean but not bare-bones. The Web UI and VS Code sidebar reuse the same chat surface. Capabilities & Tools A complete capability set is pre-built: LLM can execute operation scripts to control the computer and send reactions proactively; common Skills are built-in; full image-to-text, native PDF and Office reading, and image generation are supported; tool modifications are reversible; tool execution and code changes can undergo multi-angle AI review. API provider onboarding is streamlined and ready to use. LLM can autonomously manage MCP, skills, personas, and the organization. Memory & Context Long conversations are dynamically compressed and archived; a single session can persist indefinitely, with context staying effective through continuous compression and organization. The memory system is low-cost and comprehensive — the more you use it, the better the AI understands you. Engineering & Reliability High performance, concurrent, fast to respond. Local sessions support message delivery, session branching, goals, scheduled tasks, and manual or autonomous delegation; remote sessions support sending and receiving files and images. Tool execution has a review chain; code changes can be validated from multiple angles. Real Usage Scenarios The following are not hypotheticals — they actually happened: Since v0.8, PAI has been used to develop PAI itself Users have been using PAI continuously for financial analysis and news monitoring Users have been using PAI via WeChat remote contacts to produce Xiaohongshu content, with over a thousand published posts Users have been using PAI to analyze research papers and published papers based on it Users have been using PAI for scheduled web scraping A user ran PAI continuously for 20 hours on a programming task — it reviewed, resolved, researched online, and passed on its own Users have been using PAI long-term to create game guides Users have been using PAI long-term to operate games and complete daily tasks Users run dozens of sessions simultaneously, using PAI to monitor multiple online channels at once After extended use, users consistently report it gets smoother over time — the AI understands them better Current Shape Continuously self-iterating; current release is on the 0.97 line Desktop app plus Web UI and VS Code sidebar Local sessions, remote sessions, memory, review, delegation, tasks, multi-window, and workspace capabilities all shipped Tech Stack Desktop shell: Tauri 2 Backend: Rust (async, tokio) Frontend: Vue 3 + TypeScript + Vite UI: DaisyUI + Tailwind CSS Package manager: pnpm Platform & Updates Current release strategy: Windows: NSIS installer + zip portable (PORTABLE marker), in-app auto-update Linux: .deb / AppImage, release pipeline maintained macOS: universal DMG for Intel and Apple Silicon, release pipeline maintained In-app auto-update currently covers Windows installer and portable builds only. Data & Privacy API keys are stored locally, never passing through any intermediate server Conversations, tasks, archives, memory, and media are all stored locally Standard install uses the OS config directory; portable data lives in data/ next to the executable You can manage, export, and clean up all your data yourself Who It Is For Developers who want AI truly embedded in their desktop workflow People not satisfied with AI tools that "can only chat" People who need long-running task execution, not one-shot Q&A People who want AI with review capability, not blind delegation People with imagination for AI organizational collaboration Quick Start Download the installer, portable zip, Linux package, or macOS DMG from Releases. Default data locations: Windows: %APPDATA%\easycall\p-ai Linux: ~/.config/p-ai macOS: ~/Library/Application Support/ai.easycall.p-ai Portable: data/ next to the executable On Linux, the AUR package installs as easy-call-ai; official .deb / AppImage builds follow the release artifacts. Documentation Document index — current implementation notes, historical drafts, and which docs should still be treated as source of truth AGENTS.md — project conventions for coding agents Acknowledgments This project relies on these excellent upstream projects and communities: Tauri · Vue 3 · DaisyUI · Tailwind CSS · rust-genai · rmcp · Shiki · Mermaid · KaTeX · tokio · reqwest · rusqlite ·