Fabro, a Graphviz-Based AI Orchestrator, Released

Qlty Software has released Fabro, a coding agent orchestrator, under the MIT license. Fabro uses Graphviz DOT syntax to define agents, shell commands, and human approval procedures as graphs, enabling them to be version-controlled and reviewed like source code.

Key Features and How It Works

• Graphviz-Based Procedure Definition: Write graph structures in .fabro files to explicitly control conditional branching, loops, parallel execution, and Human Gates.
• Per-Step Model Assignment: Use the model_stylesheet attribute to specify the LLM model and reasoning effort for each node via CSS-like rules. For example, optimize token costs by using low-cost models for planning steps and frontier models for implementation steps.
• Various Node Types: Supports 7 node types: Agent, Prompt, Command, Human, Wait, Conditional, and Parallel.
• Execution History and Verification: Records inputs, outputs, and token usage for every execution Stage, and automatically performs a correction loop upon verification failure. Changes are committed to a Git branch for traceability.

Deployment and Environment

• Single Executable: Developed in Rust with no runtime dependencies, allowing usage without installing Python or Node.js.
• Installation Methods: Can be installed via 4 methods: Claude Code, Codex, Homebrew, and Bash script. Currently, GitHub releases offer only -nightly prereleases.
• Security and Hosting: Distributed with SLSA build provenance, supporting local mode or self-hosting via Docker/Kubernetes, etc.

Suitability

Suitable for small specialized development teams that need to share repetitive development procedures and require automatic correction loops when verification steps fail. Conversely, it is unsuitable for exploratory tasks with undefined procedures or one-off tasks, as the cost of graph design may be excessive.