Open Source 开源项目 6h ago Updated 6h ago 更新于 6小时前 48

GitHub: askimo-ai/askimo GitHub:askimo-ai/askimo

Askimo is an open-source, native desktop AI client (AGPLv3) built with Kotlin Compose Multiplatform that aggregates multiple AI providers into a single application It supports both cloud APIs (OpenAI, Claude, Gemini, Grok, etc.) and local inference via Ollama, LM Studio, vLLM, and Docker AI, with per-session provider switching Key differentiators include local RAG with hybrid BM25 + vector retrieval, multi-step agentic workflow plans (YAML or form-based), MCP tool integration, and an integrated Askimo是一款基于Kotlin Compose Multiplatform开发的跨平台桌面AI客户端,支持OpenAI、Claude、Gemini、Grok等云端模型及Ollama、LM Studio等本地模型 核心亮点是本地RAG检索增强生成,用户文件完全保留在本地设备,通过BM25+向量混合检索实现隐私优先的文档搜索与对话 内置多步骤AI工作流引擎(Plans)、脚本执行器(Python/Bash/JS)和MCP工具集成,支持从表单到PDF/Word的完整交付链 采用SQLite本地持久化会话存储,支持长对话不丢失上下文,内存占用仅50-300MB,磁盘空间250MB 开源协议为AGP

55
Hot 热度
52
Quality 质量
50
Impact 影响力

Analysis 深度分析

TL;DR

  • Askimo is an open-source, native desktop AI client (AGPLv3) built with Kotlin Compose Multiplatform that aggregates multiple AI providers into a single application
  • It supports both cloud APIs (OpenAI, Claude, Gemini, Grok, etc.) and local inference via Ollama, LM Studio, vLLM, and Docker AI, with per-session provider switching
  • Key differentiators include local RAG with hybrid BM25 + vector retrieval, multi-step agentic workflow plans (YAML or form-based), MCP tool integration, and an integrated script runner for Python/Bash/JS
  • All data stays on-device: local SQLite session storage, local telemetry, and no cloud upload of user files or conversations
  • Built on JDK 25+, requiring 50–300 MB RAM and 250 MB disk, with cross-platform support for macOS 11+, Windows 10+, and major Linux distributions

Why It Matters

Askimo represents a growing trend toward unified, privacy-first AI desktop clients that eliminate the friction of managing multiple provider dashboards and browser tabs. For AI practitioners who work with sensitive data or require hybrid cloud/local model setups, it offers a practical, self-hostable solution with agentic workflow capabilities typically found only in enterprise platforms.

Technical Details

  • Architecture: Kotlin Compose Multiplatform desktop app with a modular structure (desktop/, desktop-shared/, shared/, cli/); core logic in shared/ handles providers, RAG, MCP, memory, tools, database, plans engine, and agent runtimes
  • RAG Pipeline: Hybrid BM25 + vector retrieval with an AI classifier that conditionally skips retrieval when the query does not require it; indexes local folders, files, and web URLs with data remaining entirely on-device
  • Agentic Workflows: "Plans" engine allows chaining multi-step AI pipelines from a form UI or YAML definitions; each step consumes prior output, with live progress and PDF/Word export
  • Script Runner: Executes Python (in auto-managed virtualenv with pip install), Bash, and JavaScript directly from chat; MCP tools connect via stdio or HTTP
  • Provider Abstraction: Unified interface for OpenAI-compatible endpoints, supporting OpenAI, Anthropic Claude, Google Gemini, xAI Grok, OpenRouter, NVIDIA NIM, Together AI, Ollama, LM Studio, Docker AI, vLLM Server, and custom base URLs
  • Persistence & Telemetry: Conversations stored in local SQLite; local telemetry tracks token usage, cost estimates, and RAG performance per provider with no external data transmission

Industry Insight

  • The consolidation of multi-provider chat, local RAG, agentic workflows, and MCP tooling into a single offline-capable desktop app signals that end-users increasingly demand privacy-preserving AI workstations rather than browser-based portals; expect continued growth in the "local-first AI client" category.
  • The inclusion of a plans engine with YAML-defined workflows and live step chaining positions Askimo closer to lightweight agentic platforms, suggesting that the barrier between consumer AI chat clients and professional agent tooling is narrowing.
  • For organizations handling sensitive data, Askimo's local-only RAG and telemetry model offers a ready-made compliance-friendly alternative to cloud-dependent AI tools, particularly valuable in regulated industries where data residency is mandatory.

TL;DR

  • Askimo是一款基于Kotlin Compose Multiplatform开发的跨平台桌面AI客户端,支持OpenAI、Claude、Gemini、Grok等云端模型及Ollama、LM Studio等本地模型
  • 核心亮点是本地RAG检索增强生成,用户文件完全保留在本地设备,通过BM25+向量混合检索实现隐私优先的文档搜索与对话
  • 内置多步骤AI工作流引擎(Plans)、脚本执行器(Python/Bash/JS)和MCP工具集成,支持从表单到PDF/Word的完整交付链
  • 采用SQLite本地持久化会话存储,支持长对话不丢失上下文,内存占用仅50-300MB,磁盘空间250MB
  • 开源协议为AGPLv3,支持macOS 11+、Windows 10+、Linux多发行版,提供9种语言本地化

为什么值得看

Askimo代表了AI客户端工具从"云端依赖"向"本地优先+隐私可控"演进的重要趋势,为开发者提供了可参考的桌面端AI应用架构范式。其将RAG、Agent工作流、MCP协议集成于单一原生应用的设计思路,对构建企业级AI工具链具有直接借鉴价值。

技术解析

  • 架构设计:采用Kotlin Compose Multiplatform构建原生桌面应用,分为desktop(UI层)、desktop-shared(共享组件)、shared(核心逻辑层:Provider、RAG、MCP、Memory、Tools、Database、Plans Engine)三层架构,通过Gradle构建系统管理
  • RAG实现:本地文件索引采用BM25关键词检索与向量检索混合策略,内置AI分类器智能判断是否需要检索,避免无效查询;数据全程本地处理,不上传至任何云服务
  • 工作流引擎:Plans系统支持通过表单UI或YAML定义多步骤AI管道,每步依赖上一步输出,实时展示进度,支持PDF/Word导出;Script Runner提供自动管理的Python虚拟环境,支持依赖自动安装
  • 模型接入:支持OpenAI兼容API的所有提供商(OpenRouter、NVIDIA NIM、Together AI、vLLM等),本地模型通过Ollama/LM Studio/Docker接入,Web搜索集成DuckDuckGo、Brave、Tavily、SearxNG等多后端
  • 工程特性:会话数据本地SQLite存储实现持久化,本地遥测系统记录Token用量、成本估算、RAG性能指标,Vision功能支持多模态图像输入,MCP通过stdio/HTTP协议接入

行业启示

  • 本地优先架构成为差异化竞争点:在云端API同质化严重的背景下,隐私保护和本地处理能力正成为AI客户端的核心竞争力,Askimo的"数据不出设备"设计契合企业合规需求
  • MCP协议生态加速工具链整合:通过标准协议接入任意MCP服务器,降低了AI应用与外部工具集成的开发成本,预示未来AI客户端将演变为"工具聚合平台"而非单一聊天界面
  • 原生桌面应用体验优势凸显:相比Web Wrapper方案,原生应用在长会话稳定性、资源占用、系统级集成(文件搜索、脚本执行)方面具有显著优势,适合重度生产力场景

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Code Generation 代码生成 Programming 编程 Deployment 部署