What it does
LingShu is an Apache-2.0 execution agent that runs as a Mac app, with a Windows technical preview. You bring the model: the site names OpenAI, Anthropic, DeepSeek and MiniMax, or a compatible custom endpoint. It places itself in the same category as Codex and Claude Code but treats code as only one of several deliverables. A request is turned into a structured goal with acceptance criteria. A main agent plans the work, passes pieces to isolated worker sessions and local tools, and sends the result to an independent checker before calling it complete. For reports and slide decks it writes real .pptx, .docx or other files to disk, registers them in an artifact ledger, previews them in the app and revises them. The README admits that in its public sample the first checker missed defects that an external render exposed. On macOS it can read accessibility snapshots and operate apps once you authorize it, and task records, artifacts and memory stay on your computer. A bundled lingshu CLI and a local HTTP JSON-RPC control plane are included. The README labels the project alpha-stage, advises reviewing permissions and keeping backups, and says direct computer control on Windows is intentionally unavailable.
Capabilities
- Model-agnostic: OpenAI, Anthropic, DeepSeek, MiniMax or a compatible endpoint
- Creates real code, PPTX, DOCX, PDF and Markdown files that are previewed and checked
- Isolated worker sessions with an independent checker before completion
- Native macOS computer use through accessibility snapshots, after authorization
- Task records, artifacts and memory kept locally
- Bundled lingshu CLI and a local HTTP JSON-RPC control plane
When to use it
- Producing a finished slide deck or report from a brief rather than just an outline
- Delegating Mac workflows to an agent running on a model you pick
- Inspecting or modifying the full source of an agent runtime
Alternatives to explore
- OpenAI Codex
- Claude Code