Bring your own model
自带主脑
OpenAI, Anthropic, DeepSeek, MiniMax, and compatible endpoints share the same agent layer.
OpenAI、Anthropic、DeepSeek、MiniMax 与兼容端点共用同一能力层。
Open-source, model-agnostic macOS agent
纯开源、模型无关的 macOS Agent
LingShu
Your model. Your agent runtime. More than code.
你的模型，你的 Agent；代码不是边界。
Use OpenAI, Claude, DeepSeek, MiniMax, or a compatible endpoint. Deliver verified software, editable presentations and documents, and authorized Mac workflows through one Apache-2.0 runtime.
支持 OpenAI、Claude、DeepSeek、MiniMax 与兼容端点。通过同一套 Apache-2.0 运行时，交付经过验收的软件、可编辑 PPT/文档与授权后的 Mac 工作流。
- macOS 14+
- Model-agnostic模型无关
- Apache-2.0
- Apple notarizedApple 已公证
Full runtime open source完整运行时开源Native app + agent layer原生 App + Agent 能力层
Bring your own model自带主脑No single-vendor lock-in不锁定单一模型厂商
Beyond code不只交付代码PPTX, DOCX, PDF, media, appsPPTX、DOCX、PDF、媒体、应用
Verified delivery验收式交付Artifacts, preview, checker产物、预览、独立 checker
The execution-agent category
执行型 Agent 赛道
Same category. Different boundaries.同一个赛道，不同的边界。
Codex and Claude Code set the standard for execution-oriented coding agents. LingShu starts from the same agent premise, then makes the complete native runtime open source, keeps the model replaceable, and treats code as one deliverable among several.
Codex 与 Claude Code 定义了执行型编程 Agent 的标杆。灵枢从同一类 Agent 出发，把完整原生运行时开源、把主脑做成可替换能力，并让代码成为多种交付物之一。
Dimension维度
LingShu
OpenAI Codex
Claude Code
Product focus产品定位
General execution on macOSmacOS 通用执行
Software engineering软件工程
Software engineering软件工程
Open-source surface公开开源范围
Native app + runtime, Apache-2.0原生 App + 运行时，Apache-2.0
Codex CLI, Apache-2.0Codex CLI，Apache-2.0
Official repository: all rights reserved官方仓库：All Rights Reserved
Model layer模型层
OpenAI, Claude, DeepSeek, MiniMax, compatible endpointsOpenAI、Claude、DeepSeek、MiniMax、兼容端点
OpenAI modelsOpenAI 模型
Claude modelsClaude 模型
Default output默认交付范围
Code, office files, media, Mac actions代码、Office 文件、媒体、Mac 操作
Code and engineering work代码与工程任务
Code and engineering work代码与工程任务
Positioning, not a benchmark. Sources: official Codex repository, Codex product page, Claude Code repository, and Claude Code product page. Product surfaces and model capabilities change over time.
这是品类与公开形态对比，不是性能榜。依据：官方 Codex 仓库、Codex 产品页、Claude Code 仓库与Claude Code 产品页。产品形态与模型能力会持续变化。
Execution-agent workflow
执行型 Agent 工作流
From a goal to a checked file从目标到经过验收的文件
The main conversation protects intent. Isolated workers do the long work. Every deliverable returns to one traceable record.
主会话负责守住意图，隔离子任务推进长流程，所有产物最终回到同一条可追踪任务记录。
- 01Understand理解
Turn the brief into a stable goal and acceptance boundary.
把需求转成稳定目标与验收边界。
- 02Plan规划
Build a runtime workflow that can adapt without changing the goal.
构建可动态调整、但不篡改目标的运行图。
- 03Execute执行
Dispatch workers, tools, local files, and authorized Mac apps.
调度子任务、工具、本地文件与已授权的 Mac 应用。
- 04Collaborate协作
Pause for login, QR, file choice, or confirmation and resume exactly.
遇到登录、扫码、选文件或确认时精确暂停并续接。
- 05Verify验收
Register the real artifact and send it to an independent checker.
登记真实产物，并交给独立 checker 验收。
Three-minute path
三分钟上手路径
Run your first traceable task运行第一条可追踪任务
Install LingShu, choose a language, connect your own model, then start with a small local deliverable.
安装灵枢、选择语言、连接自己的主脑，然后从一个小型本地产物开始。
macOS 14+OpenAIClaudeDeepSeekMiniMax M3
Recommended: Homebrew推荐：Homebrew
brew install --cask RoyZhao1991/tap/lingshu
Try this first
第一条任务可以这样写
Create a one-page project brief in DOCX, save it locally, preview it, and have an independent checker verify the result.
生成一份单页项目简报 DOCX，保存到本地，打开预览，并交给独立 checker 验收。Share a 15-minute first-run report提交 15 分钟首跑报告
Flagship capability
旗舰能力
Complete reports, not outlines交付完整报告，而不是只给提纲
LingShu can take a presentation or document brief through narrative design, file generation, preview, revision, artifact registration, and independent review.
灵枢可以把演示或文档需求一路推进到叙事设计、文件生成、预览修改、产物登记与独立复核。
- Editable PPTX and DOCX output可编辑的 PPTX 与 DOCX
- Built-in local preview and artifact ledger内置本地预览与产物清单
- Independent checker before completion完成前进行独立验收
- Presentation narration and content Q&A演示讲解与基于内容的问答
Public reproducible sample
公开可复现样例
Inspect the files, not a claim直接检查文件，而不是相信一句“已完成”
A single fictional Project Aurora brief became an editable four-slide PPTX and a one-page DOCX. Independent rendering exposed defects, and LingShu revised the same goal until the final deck passed structural and visual checks.
一条虚构的 Project Aurora 需求生成了四页可编辑 PPTX 和一页 DOCX。独立渲染发现初稿问题后，灵枢围绕同一目标持续修订，直到最终文件通过结构与视觉检查。
The public run log documents what the first checker missed and why the deck was revised. Alpha means showing the limits as well as the result.
公开记录同时说明第一轮 checker 漏掉了什么、为何继续修订。Alpha 阶段既展示结果，也如实展示边界。
A useful agent needs boundaries
能做事，也要有边界
Local by design, explicit when it matters本地优先，关键动作明确确认
Credentials stay in Keychain
凭据保存在钥匙串
Provider tokens are runtime configuration and never belong in the repository.
模型 Token 属于运行时配置，不进入代码仓库。
Permission-scoped control
受权限约束的控制
Accessibility, screen, microphone, and camera access are requested only for chosen capabilities.
仅在使用相应能力时请求辅助功能、屏幕、麦克风与摄像头权限。
Typed human checkpoints
结构化人机协作
Login, QR, selection, confirmation, and physical actions pause the exact task instead of disappearing into terminal text.
登录、扫码、选择、确认与实体操作会暂停精确任务，而不是藏在终端文本里。
Alpha notice: review requested permissions, keep backups, and inspect important deliverables before external use.
公测提示：请核对权限请求、保留重要文件备份，并在对外使用前检查关键产物。
Built in the open
开放共建
Inspect it. Run it. Improve it.查看、运行、一起改进。
The complete native app and agent runtime are Apache-2.0 licensed, written in Swift, and ship with more than 1,500 tests. Bring your own model, inspect the orchestration, and help shape an agent layer that is not owned by one provider.
完整原生 App 与 Agent 运行时采用 Apache-2.0，以 Swift 编写，并拥有超过 1,500 项测试。你可以自带主脑、审阅完整编排，并共同塑造一套不属于单一模型厂商的 Agent 能力层。