What Sgai Does

Sgai (pronounced "Sky") is a goal-driven AI software factory that runs locally in your repository. Instead of step-by-step prompting, you define outcomes in GOAL.md files describing what should be built, not how. The system then coordinates multiple AI agents to execute the goal.

Key Features

Goal-driven workflow: Define outcomes in GOAL.md files with completion gates (e.g., make test) that determine when work is actually done
Multi-agent coordination: Decomposes goals into a DAG of roles (developer → reviewer → safety analyst, etc.)
Local execution: Everything runs locally in your repo with no automatic pushes to GitHub
Visual monitoring: Web dashboard shows real-time execution of the agent graph
Interactive clarification: Agents ask clarifying questions when needed before execution
Skill extraction: Extracts reusable skills and code snippets from completed sessions

How It Works

The workflow follows these steps:

Create a Goal: Most users create goals using the built-in wizard. Goals are stored in GOAL.md and describe outcomes, not implementation steps.
Agents Plan the Work: Sgai breaks your goal into a workflow diagram of coordinated agents with defined roles.
Approve & Monitor: Agents ask clarifying questions, then work autonomously executing tasks, running tests, and validating completion.
Learn from Sessions: The system extracts reusable skills from completed sessions.

Example GOAL.md

--- flow: | "backend-developer" -> "code-reviewer" completionGateScript: make test interactive: yes ---

Build a REST API

Create endpoints for user registration and login with JWT auth.

POST /register validates email, hashes password
POST /login returns JWT token
Tests pass before completion

Installation & Setup

Recommended automated setup via opencode:

opencode update
opencode auth login
opencode --model anthropic/claude-opus-4-6 run "install Sgai using the instructions from https://github.com/sandgardenhq/sgai/blob/main/INSTALLATION.md"

Manual installation requirements: Go, Node.js, bun, opencode. Recommended: jj (version control), tmux (session management), ripgrep (code search), Graphviz (diagram rendering).

Install command:

go install github.com/sandgardenhq/sgai/cmd/sgai@latest

Or build from source:

git clone https://github.com/sandgardenhq/sgai.git
cd sgai
cd cmd/sgai/webapp && bun install && cd ../../..
make build

Running Sgai

Start the server with sgai serve and open http://localhost:8080 to access the dashboard.

Technical Details

Open source (Go)
Works with Anthropic, OpenAI, or local models via opencode
Changes go through your version control (recommends jj, but Git works)
Demo available: 4-minute video
Example use case: "Build a drag-and-drop image compressor" → 3 agents (developer, reviewer, designer) → Working app with tests passing → 45 minutes

The project is described as "still early and rough in places, but functional enough to share" and has been used internally for prototyping small apps and internal tooling.

📖 Read the full source: HN AI Agents