Defract is an agent orchestration harness and desktop workflow environment designed to coordinate multi-stage AI software engineering workflows. Built as an operational layer above underlying coding agents, it structures AI development into a discrete pipeline covering backlog grooming, scoping, architecture, implementation, automated review, and release. In day-to-day operation, Defract combines a multi-pane desktop interface for macOS and Windows with a native CLI. Instead of executing changes sequentially in a single working directory, Defract spins up isolated Git worktrees for every task. Developers approve high-level acceptance criteria, design parameters, and system architecture, while specialized agent roles handle file modifications, type checking, test validation, and peer code reviews. The system supports full Autopilot execution for high-velocity pipelines as well as Human-in-the-Loop checkpoints for sensitive modules. Defract differentiates itself from standard terminal CLI tools by replacing manual diff micromanagement with structured agent-to-agent verification and parallel worktree management. It also features an embedded memory system that captures repository conventions, past design choices, and bug learnings across completed stories, applying compound context to future tasks. Currently in open beta, Defract is free to download for macOS and Windows. It natively runs Claude Code out of the box, with roadmap support planned for OpenAI Codex, Google Gemini, and OpenCode.
Isolated Git Worktrees — Automatically provisions a dedicated Git worktree per story to enable concurrent agent execution without branch conflicts or dirty repository states. - Six-Stage Development Pipeline — Organizes agent activity through structured phases: Scoping, Design, Architecture, Implementation, Review, and Release. - Deep Claude Code Protocol Integration — Interacts directly with Claude Code's protocol for low-overhead context retrieval and tool access. - Agent-on-Agent Code Review — Runs automated agent verification stages for type safety, test execution, and acceptance criteria matching before code merges. - Dual Autopilot and Supervised Modes — Allows toggling between an automated Autopilot mode for parallel feature delivery and a Human-in-the-Loop mode for step-by-step gate reviews. - Compounding Repository Memory — Persists codebase conventions, past architectural decisions, and contextual nuances to refine downstream task accuracy.
Use cases
Use Case: Parallel Multi-Story Development Running multiple distinct feature or bugfix tasks concurrently across separate Git worktrees, allowing agents to groom, implement, and test separate stories without workspace collisions.
Use Case: Structured Issue Scoping and Backlog Grooming Transforming rough task descriptions or feature requests into fully structured specifications with automated criteria, edge-case analysis, and risk checks before any code is modified.
Use Case: Automated Agent-Gated Verification Delegating implementation and peer review to specialized agent pipelines that verify test suites and typing before passing merge sign-offs, freeing developers from manual diff babysitting.
Who it is for
Defract is designed for fast-shipping software engineers and AI-native engineering teams that need to scale beyond single-agent terminal chats into parallel, multi-task workflows. It is not suitable for no-code users, air-gapped repositories without API connectivity, or teams seeking fully unsupervised merging into main without initial architectural review.
Install guide
Installation Download the Defract desktop installer directly from the official website for your operating system: - macOS: Download the .dmg installer from https://defract.dev/download - Windows: Download the .exe installer from https://defract.dev/download Ensure Claude Code CLI is installed on your system path: bash npm install -g @anthropic-ai/claude-code
First Run 1. Launch the Defract application and authenticate your Anthropic API credentials or Claude Code session. 2. Open an existing Git repository inside Defract to allow the harness to initialize repository indexers and Git worktree orchestration. 3. Configure your desired operating mode (Autopilot or Human-in-the-Loop) under project settings.
Example Session To create and scope a new story from the integrated terminal or Defract prompt interface: bash # Initialize a new scoped task story new "add stripe checkout to settings page" # Inspect generated criteria and run automated parallel execution story queue --auto
Superset — Superset is an open-source agent orchestration harness and multi-agent workspace developed by the superset-sh team. It provides a structured environment…
Okto Nexus — Okto Nexus is a local-first multi-agent coordination harness and orchestrator developed by OktoLabs. It acts as an execution bus designed…
Conductor — Conductor is a multi-agent orchestration harness developed by Melty Labs. It is designed to manage, coordinate, and execute parallel coding…
Bwee — Bwee is a native macOS agent harness designed to organize terminal-based AI agents and render interactive user interfaces on the…
Ona — Ona (formerly Gitpod's agentic platform, now part of OpenAI) is an enterprise-grade cloud agent orchestration platform that manages and runs…
Kiro — Kiro is an agentic engineering platform and CLI coding agent built and operated by AWS. It extends beyond standard code-completion…