Sandgarden has open-sourced Sgai, a goal-driven AI development tool that builds software by running coordinated agent workflows. Instead of step-by-step prompting, developers define outcomes in a GOAL.md file, and Sgai decomposes the task into a directed acyclic graph of specialized roles like developer, reviewer, and safety analyst.

Built in Go with a TypeScript frontend, Sgai runs entirely locally in your repository with a web dashboard showing real-time execution. The system uses completion gates—such as requiring tests to pass—before marking work as done. Users can supervise progress, answer clarifying questions, and approve plans before agents work autonomously. The team has used it internally for prototyping apps and internal tooling.

The project supports Anthropic, OpenAI, or local models via opencode and includes automated setup through opencode scripts. With 26 GitHub stars and a 4-minute demo video, Sgai represents a novel approach to AI-assisted development that emphasizes transparency and user control over hidden AI reasoning.