helixml/helix

What it solves

Helix provides a way to run a fleet of coding agents in parallel on your own infrastructure, moving away from the model of a single agent running in a local terminal. It allows teams to collaboratively manage AI coding tasks via a shared Kanban board, ensuring that agents have isolated, GPU-accelerated environments to work in without competing for resources on a developer's laptop.

How it works

Helix organizes work into projects with a Kanban board where tasks move through a lifecycle: Backlog, Planning, Spec Review, In Progress, and Pull Request. A planning agent first creates a technical specification; once approved, an implementation agent executes the code changes in an isolated sandbox. Each agent is provided with a full GPU-accelerated streaming desktop (including browser, terminal, and GUI apps) rather than just a shell. The system is designed to be harness-agnostic, supporting various agent frameworks (like Claude Code or Gemini CLI) and any OpenAI-compatible LLM provider, including self-hosted models via vLLM.

Who it’s for

It is designed for engineering teams who want to run multiple AI coding agents on their own servers (including air-gapped environments) and collaboratively steer and review their work through a pull request workflow.

Highlights

• Isolated Desktops: Each agent gets its own GPU-accelerated desktop with a full GUI, browser, and filesystem.
• Spec-Driven Workflow: Tasks require a planning phase and a spec review before implementation begins.
• Fleet Visibility: A central dashboard allows users to watch agents work in real-time and jump in for pair programming.
• Harness and Model Agnostic: Supports multiple agent harnesses (ACP-compatible) and any OpenAI-compatible LLM provider.
• Multiplayer Collaboration: Shared agent environments persist state and chat history for team handoffs.
• Enterprise Ready: Includes built-in RAG, OIDC SSO, role-based access control, and token usage metering.