LingShu
GitHub

Architecture, not a model wrapper

这不是一层模型外壳

The brain is replaceable.
The runtime is the product.主脑可以替换，
运行时才是产品。

LingShu keeps model intelligence behind an adapter and puts durable agency in an Apache-2.0 native macOS runtime: goal preservation, mutable workflows, isolated workers, human checkpoints, artifact registration, memory, and independent verification.

灵枢把模型智能放在适配器之后，把持久的 Agent 能力放进 Apache-2.0 原生 macOS 运行时：目标守恒、动态工作流、隔离子任务、人机检查点、产物登记、记忆与独立验收。

Inspect the runtime查看运行图 Read the source查看源码
  • Apache-2.0complete app + runtime完整 App + 运行时
  • 5execution stages个执行阶段
  • 1,581release tests项发布测试
  • macOS 14+native Swift app原生 Swift App

One open runtime, many brains

一套开放运行时，多种主脑

The model proposes. The runtime preserves, executes, and verifies.模型负责提出判断，运行时负责守住目标、执行与验收。

A provider can change without replacing the task ledger, permission boundary, local tools, memory, or delivery contract.

切换模型服务商，不需要替换任务账本、权限边界、本地工具、记忆或交付契约。

What the model cannot silently rewrite

模型不能静默改写什么

Agency needs contracts, not a longer prompt.Agent 需要的是契约，不是更长的提示词。

01

Goal stays fixed

目标保持不变

If a valid GoalSpec cannot be generated, the task does not begin with a weaker substitute.

如果不能生成完整 GoalSpec，就不以更弱的默认目标启动任务。

02

Plans may change

计划可以调整

The brain may revise only pending steps as evidence and tool results arrive.

随着证据和工具结果到达，主脑只能调整尚未执行的步骤。

03

Main context stays serialized

主上下文保持串行

Long work runs in isolated child sessions; distilled completion memory returns to the main conversation.

长任务在隔离子会话中运行，完成记忆再显式注入主会话。

04

Human action is typed

人机交互是结构化状态

A login, QR scan, file choice, physical step, or approval pauses the exact session and resumes it later.

登录、扫码、选文件、实体操作或授权会暂停精确会话，并从原处续接。

05

Artifacts must exist

产物必须真实存在

The registry is built from actual file and Git changes, not from names claimed in model text.

产物清单来自真实文件与 Git 变化，而不是模型文字中声称的文件名。

06

Verification is independent

验收必须独立

A separate checker evaluates the delivery against the original acceptance boundary.

独立 checker 按原始验收边界检查最终交付。

Same execution-agent category

同属执行型 Agent 赛道

LingShu changes the ownership boundary.灵枢改变的是能力归属边界。

Codex and Claude Code are the relevant comparison, not chat applications. LingShu makes a different product choice: users can inspect and modify the complete native runtime while choosing the model behind it.

相关对比对象是 Codex 与 Claude Code，而不是聊天应用。灵枢做了不同的产品选择：用户既能审阅和修改完整原生运行时，也能自行选择背后的模型。

Open surface
开源范围
Complete Swift app and agent runtime under Apache-2.0.
完整 Swift App 与 Agent 运行时采用 Apache-2.0。
Model ownership
模型选择权
Bring OpenAI, Claude, DeepSeek, MiniMax, or a compatible endpoint.
可接入 OpenAI、Claude、DeepSeek、MiniMax 或兼容端点。
Delivery boundary
交付边界
Code, editable office files, media, and authorized Mac actions share one ledger.
代码、可编辑 Office 文件、媒体与授权 Mac 操作共用同一任务账本。
Proof model
验收方式
Files are registered, previewed, and independently checked before completion.
文件在完成前会被登记、预览并接受独立验收。

This is an architectural positioning statement, not a performance benchmark. Product surfaces change. See the official Codex repository and Claude Code repository.

这是架构定位说明，不是性能榜。产品形态会持续变化，参见官方 Codex 仓库与 Claude Code 仓库。

Do not take the diagram on faith

不要只相信这张图

Inspect the code, run a sample, open the artifacts.检查源码、复跑样例、打开真实产物。

01Architecture document总体架构文档System boundaries and module map系统边界与模块地图 02Project Aurora sampleProject Aurora 样例Editable PPTX, DOCX, PDF, and rendered slides可编辑 PPTX、DOCX、PDF 与渲染页 03Real task replay真实任务回放Goal, artifact ledger, and checker result目标、产物账本与验收结果 04Build and test history构建与测试记录Public CI for the open runtime开放运行时的公开 CI

Bring your model. Keep the agent layer.带上你的模型，保留你的 Agent 能力层。

Install the signed Alpha or read the source before trusting the claim.

安装签名公测版，或先读源码再判断。

Download signed Alpha下载签名公测版 Star on GitHub在 GitHub 点 Star