# LingShu

> LingShu is an Apache-2.0, model-agnostic native macOS execution agent in the Codex / Claude Code category. It uses a replaceable model backend and treats code, editable presentations, documents, media, and authorized Mac actions as first-class deliverables.

## Official Links

- Website and real task replay: https://royzhao1991.github.io/LingShu/
- Shareable architecture and value map: https://royzhao1991.github.io/LingShu/architecture/
- Source repository: https://github.com/RoyZhao1991/LingShu
- Signed macOS Alpha: https://github.com/RoyZhao1991/LingShu/releases/tag/v0.1.0-alpha.8
- Architecture: https://github.com/RoyZhao1991/LingShu/blob/main/Docs/ARCHITECTURE.md
- Chinese README: https://github.com/RoyZhao1991/LingShu/blob/main/README.zh-CN.md
- Discussions: https://github.com/RoyZhao1991/LingShu/discussions

## Verified Project Facts

- The complete Swift native app and agent runtime are published under Apache-2.0.
- Users can configure OpenAI, Anthropic Claude, DeepSeek, MiniMax, or a compatible endpoint.
- The runtime includes goal structuring, mutable execution workflows, isolated child tasks, local tools, permission gates, task records, artifact registration, human checkpoints, memory, and independent verification.
- A public Project Aurora example includes a real editable PPTX, DOCX, PDF, rendered slides, and a task replay assembled from the corresponding task record.
- The public Universal DMG supports Apple silicon and Intel, is signed with Apple Developer ID, notarized, and distributed with a SHA-256 checksum.
- Homebrew install: `brew install --cask RoyZhao1991/tap/lingshu`.

## Important Boundaries

- LingShu is Alpha software for macOS 14 or later, not a claim of unattended autonomy.
- Users provide their own supported model token or compatible endpoint.
- The app, orchestration, task records, and artifacts run locally, but selected context is sent to the model provider configured by the user.
- Model behavior, latency, multimodal support, and output quality vary by provider.
- Important computer actions and generated deliverables should be reviewed.
- Positioning references to Codex and Claude Code identify the execution-agent category; they are not benchmark or superiority claims.

## Reproducible Evidence

- Project Aurora run and editable files: https://github.com/RoyZhao1991/LingShu/tree/main/Examples/project-aurora
- Architecture field guide: https://github.com/RoyZhao1991/LingShu/blob/main/Docs/ARCHITECTURE.md
- Security boundaries: https://github.com/RoyZhao1991/LingShu/blob/main/SECURITY.md
- Changelog: https://github.com/RoyZhao1991/LingShu/blob/main/CHANGELOG.md
- CI: https://github.com/RoyZhao1991/LingShu/actions/workflows/ci.yml