open-koi

# Open Koi **Leap beyond your workflow.** Open Koi is an open-source, local-first, extensible AI operating entry for agents, workflows, tools, MCP, knowledge, skills, and plugins. ## Project Structure ```text doc-en/ English product and roadmap documentation doc-zh/ Chinese product and roadmap documentation website/ Static official website ``` ## Roadmap - V0.1 CLI Core - V0.2 Open Koi Desktop Workbench - V0.3 Lightweight Workflow - V0.4 Local Knowledge + Skills - V0.5 Local Plugin Runtime - V0.6 MCP Server + Online Registry - V0.7 Agent Teams - V1.0 Open Koi Mobile ## Documentation - [中文文档](./doc-zh/README.md) - [English Docs](./doc-en/README.md) ## V0.1 CLI Core The V0.1 Rust CLI lives in `crates/` and provides the first local Agent loop, Tool Registry, sessions, audit logs, worktree commands, and stdio MCP discovery. Run it locally: ```bash cargo run -p koi-cli -- config doctor cargo run -p koi-cli -- chat --message "hello" --provider mock cargo run -p koi-cli -- tools list cargo run -p koi-cli -- sessions list cargo run -p koi-cli -- audit show ``` Use `OPEN_KOI_HOME=/path/to/home` to isolate local runtime data during testing. ## Website The static official website lives in [website/](./website/). Run it locally: ```bash python3 -m http.server 4173 ``` Then open: - Chinese: `http://127.0.0.1:4173/website/` - English: `http://127.0.0.1:4173/website/en/` For sharing publicly, deploy the website as a static site. Good options include GitHub Pages, Cloudflare Pages, Netlify, or Vercel. For GitHub Pages, `website/` is the source directory in this repository. The workflow builds a `public/` artifact from `website/`, `doc-zh/`, and `doc-en/`, then publishes that artifact to the `gh-pages` branch. This repository includes a GitHub Pages workflow at `.github/workflows/deploy-website.yml`. It publishes the website to the `gh-pages` branch on every push to `main` and includes `doc-zh/` and `doc-en/` in the public site. In GitHub repository settings, enable Pages with: - Source: `Deploy from a branch` - Branch: `gh-pages` - Folder: `/ (root)` Expected GitHub Pages URL: - `https://fengmingdev.github.io/open-koi/` - `https://fengmingdev.github.io/open-koi/en/` ## 借鉴与致谢 Open Koi 充分消化已有开源成果，按需整合而非重复造轮子。 | 项目 | 仓库 | 技术栈 | 协议 | 借鉴要点 | |------|------|--------|------|---------| | **Claw Code** | [instructkr/claw-code](https://github.com/instructkr/claw-code) | Rust | 待确认 | Agent harness 的 Rust 实现范本、sub-agent 调度、6 种 MCP transport | | **OpenClaw** | [openclaw/openclaw](https://github.com/openclaw/openclaw) | TypeScript | MIT | Channel-Agent-Tool 三层解耦、消息平台桥接思想 | | **Goose** | [block/goose](https://github.com/block/goose) | Rust | Apache-2.0 | Agent loop、Provider 抽象、MCP client、工具系统 | | **OpenCode** | [opencode-ai/opencode](https://github.com/opencode-ai/opencode) → [charmbracelet/crush](https://github.com/charmbracelet/crush) | Go | MIT | 客户端-服务器分离模型、多 Provider 抽象、SQLite 会话存储 | | **CodePilot** | [op7418/CodePilot](https://github.com/op7418/CodePilot) | TypeScript / Electron | BSL-1.1 | 多模型桌面交互模式、会话控制（暂停/恢复/回退）、远程桥接、生成式 UI | | **Claude Code UI** | [siteboon/claudecodeui](https://github.com/siteboon/claudecodeui) | TypeScript | AGPL-3.0 | Web + 移动端 GUI、远程会话管理 | | **Claude Code from Scratch** | [Windy3f3f3f3f](https://github.com/Windy3f3f3f3f/claude-code-from-scratch) / [ConardLi](https://github.com/ConardLi/easy-agent) / [ProjectBarks](https://github.com/ProjectBarks/gopher-code) | Python/TS/Go | MIT 等 | 从零搭建的施工参考、23 组件拆解 | | **OpenHermes** | [teknium/OpenHermes](https://github.com/teknium/OpenHermes) (推测) | Python | Apache-2.0 | 合成数据生成、质量过滤、指令格式标准化方法论 | | **rmcp** | Rust crate | Rust | MIT / Apache-2.0 | MCP client + server 的 Rust 实现，stdio/SSE transport | | **Extism** | [extism/extism](https://github.com/extism/extism) | Rust | Apache-2.0 | WASM 插件运行时、沙箱安全模型 | | **LanceDB** | [lancedb/lancedb](https://github.com/lancedb/lancedb) | Rust | Apache-2.0 | 嵌入式向量数据库、本地优先 | | **Tantivy** | [quickwit-oss/tantivy](https://github.com/quickwit-oss/tantivy) | Rust | MIT | 全文搜索引擎、本地优先 | | **Dify** | [langgenius/dify](https://github.com/langgenius/dify) | Python/TS | Apache-2.0 | 可视化工作流画布 UX 参考 | | **CrewAI** | [crewAIInc/crewAI](https://github.com/crewAIInc/crewAI) | Python | MIT | 多 Agent 编排模式参考 | | **Jan** | [janhq/jan](https://github.com/janhq/jan) | TypeScript / Tauri | AGPL-3.0 | Tauri + AI 桌面应用集成模式 | ### 核心启发 - **“一切皆 MCP”** —— 工具集成的底层标准：内置工具、Extism 插件、远程服务、工作流，均通过统一 MCP schema 注册 - **“工作流即 Agent 工具”** —— Agent 调用工作流如同调用任何工具；工作流也可将 Agent 作为节点 - **“多 Provider + 本地推理”** —— 敏感数据走本地模型（Ollama/llama.cpp），复杂推理走云端，透明切换 ## License Open Koi is licensed under the [Apache License 2.0](./LICENSE).