AI News AI资讯 1h ago Updated 1h ago 更新于 1小时前 41

Show HN: Remarc – better contextual feedback for AI agents Show HN:Remarc——为 AI 智能体提供更好的上下文反馈

Remarc is an open-source macOS application that serves as a feedback layer between users and AI coding agents, preserving rich context (text selections, screenshots, web elements, voice notes) attached to each comment It supports multiple input modalities: text selection, annotated screenshots, web page capture via Chrome extension, and on-device voice transcription (WhisperKit/Parakeet) with Crit Mode for longer reviews Remarc integrates with major AI coding agents (Claude Code, Codex, OMP, Cur Remarc是一款开源macOS工具,作为人类与编程AI代理之间的"反馈层",将评论与原始上下文(文本、截图、网页元素)绑定传递 支持多种输入方式:文本选择、截图标注、Chrome扩展网页捕获、语音输入(Crit Mode)和快速笔记 集成Claude Code、Codex、OMP、Cursor等主流AI代理,通过MCP协议实现上下文传递 数据完全本地存储,无账户、无遥测,支持Markdown/JSON导出和webhook自动化 技术栈:Swift 6 + SwiftUI/AppKit,WhisperKit/FluidAudio本地转录,Sparkle自动更新

58
Hot 热度
62
Quality 质量
55
Impact 影响力

Analysis 深度分析

TL;DR

  • Remarc is an open-source macOS application that serves as a feedback layer between users and AI coding agents, preserving rich context (text selections, screenshots, web elements, voice notes) attached to each comment
  • It supports multiple input modalities: text selection, annotated screenshots, web page capture via Chrome extension, and on-device voice transcription (WhisperKit/Parakeet) with Crit Mode for longer reviews
  • Remarc integrates with major AI coding agents (Claude Code, Codex, OMP, Cursor, Claude Desktop) via the Model Context Protocol (MCP), enabling agents to read comments with full attached context without manual prompt reconstruction
  • The app is privacy-first: all data stays on-device with no accounts or telemetry, and supports session-based organization, comment lifecycle management, Markdown/JSON export, and webhook integration

Why It Matters

Remarc addresses a critical gap in the AI coding agent workflow: the loss of contextual fidelity when humans hand off feedback to agents. By preserving exact selections, screenshots, and web context alongside comments, it eliminates the "paraphrase tax" that currently forces users to repeatedly reconstruct context for their agents. This approach could become a standard pattern for human-agent collaboration tools, especially as MCP adoption grows across the developer ecosystem.

Technical Details

  • Architecture: Native macOS app built with SwiftUI + AppKit in Swift 6, running from the menu bar; uses Sparkle for self-updating; vendored MCP server bundled into the app
  • Input modalities: Text selection (captures quote + source app), screenshot capture with annotation tools (arrows, shapes, blur, pixelation), Chrome extension capturing URL/CSS/layout/accessibility/React component data, on-device voice transcription via WhisperKit and FluidAudio with Crit Mode splitting recordings into separate comment cards
  • Agent integrations: MCP-based connections to Claude Code, Codex, OMP (via public remarc-agent-plugins repo, MIT-licensed), Cursor (configured by app), and Claude Desktop or any MCP client (manual); all clients receive identical context and MCP tools for reading comments, managing sessions, and resolving work
  • Data model: Comments organized into sessions (by review, project, or agent conversation), with lifecycle states (Open → Handed Off → In-Progress → Resolved), searchable History for deleted comments, and export as Markdown/JSON with configurable references and metadata; webhooks support automation tool integration
  • Privacy: Zero accounts, zero telemetry; all comments and screenshots stored locally on Mac unless explicitly handed to an agent or sent via user-configured webhook

Industry Insight

  • The Remarc model—preserving rich, structured context at the point of human feedback rather than requiring agents to infer it—represents a scalable pattern for human-AI collaboration tools; expect similar "context preservation layers" to emerge across design, product, and operations workflows beyond coding
  • MCP's role as the integration backbone (rather than proprietary connectors) demonstrates the practical value of open agent protocols; projects that embrace MCP-first integration will have a significant distribution advantage as the agent ecosystem fragments across vendors
  • The privacy-by-design approach (on-device storage, no telemetry, no accounts) is increasingly a competitive differentiator for developer tools, especially among enterprise and security-conscious users; this positioning could drive adoption in regulated or privacy-sensitive environments where cloud-based feedback tools face barriers

TL;DR

  • Remarc是一款开源macOS工具,作为人类与编程AI代理之间的"反馈层",将评论与原始上下文(文本、截图、网页元素)绑定传递
  • 支持多种输入方式:文本选择、截图标注、Chrome扩展网页捕获、语音输入(Crit Mode)和快速笔记
  • 集成Claude Code、Codex、OMP、Cursor等主流AI代理,通过MCP协议实现上下文传递
  • 数据完全本地存储,无账户、无遥测,支持Markdown/JSON导出和webhook自动化
  • 技术栈:Swift 6 + SwiftUI/AppKit,WhisperKit/FluidAudio本地转录,Sparkle自动更新

为什么值得看

Remarc解决了AI编程代理工作中一个关键痛点:人类反馈往往因上下文丢失而降低代理工作效率。它通过"评论+原始上下文"的绑定机制,让代理能直接基于用户看到的原始内容工作,而非依赖转述。

技术解析

  • 核心架构:macOS菜单栏常驻应用,通过MCP(Model Context Protocol)与AI代理通信,每个评论卡片携带原始选择/截图/网页上下文,代理可直接读取而非重建上下文
  • 输入方式:文本评论(保存引用和源应用)、截图捕获(支持箭头、形状、模糊、像素化标注)、Chrome扩展(捕获URL、CSS、布局、可访问性数据、React组件信息)、语音输入(macOS 26+支持Apple Speech/WhisperKit/Parakeet本地转录,Crit Mode自动拆分长录音为多条评论)
  • 代理集成:Claude Code、Codex、OMP通过公开仓库的插件连接,Cursor由应用配置,Claude Desktop等MCP客户端可手动连接,所有客户端获得相同的MCP工具集
  • 工作流管理:评论状态机(Open → Handed Off → In-Progress → Resolved),会话分组(按审查/项目/代理对话),永久收件箱,删除评论进入可搜索历史
  • 技术栈:Swift 6、SwiftUI + AppKit、Sparkle更新、WhisperKit + FluidAudio本地转录、KeyboardShortcuts库,MIT许可证开源

行业启示

  • 上下文传递是AI代理效率的关键瓶颈:Remarc的设计验证了"保留原始上下文而非转述"能显著提升代理工作质量,这一模式可推广至更多AI协作场景
  • 本地优先+隐私保护成为差异化竞争力:无账户、无遥测、数据本地存储的设计,在AI工具普遍依赖云服务的趋势下形成独特定位,适合对数据敏感的用户群体
  • MCP协议正在成为AI工具互操作的标准:Remarc通过MCP实现与多种代理的无缝集成,反映了行业向开放协议演进的趋势,开发者应关注MCP生态建设

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。

Agent Agent Open Source 开源 LLM 大模型