Runfusion/Fusion

Fusion – an AI‑driven software factory

What it is – Fusion is a developer‑tool platform that lets you describe a software task in plain English and then hands the work over to a team of AI agents. The agents plan the work, write code, review it, and finally merge it, all while you watch the progress on a visual dashboard. It’s essentially a multi‑agent orchestrator that automates the whole “idea → production code” pipeline.

How it works

1. Task description – You type a natural‑language request (e.g., “add a dark‑mode toggle to the settings panel”).
2. Planning agent – Reads the current repository, creates a PROMPT.md that lists the steps, file scope, and acceptance criteria.
3. Workflow execution – The task is run in an isolated Git worktree (fusion/<task‑id>). A configurable workflow (coding, quick‑fix, review‑heavy, etc.) drives the sequence of actions: generate code, run a review, apply changes, repeat.
4. Human oversight – You can set the planner’s oversight level (off, observe, steer, autonomous) and require explicit approval before merges or destructive actions.
5. Dashboard – A Kanban‑style board shows every task, its plan, live diffs, and review status. You can pause, nudge, or re‑prompt agents at any step.
6. Multi‑node sync – The same fleet of agents can be accessed from laptops, servers, or phones; state is stored in an embedded PostgreSQL database (SQLite is only for one‑time migrations).
7. Model flexibility – Supports any LLM provider (Anthropic, OpenAI, Ollama, Google Gemini, Z.ai, Kimi, etc.) and even custom local runtimes. Different models can be assigned to different workflow lanes.
8. Agent ecosystem – Comes with a library of 440+ pre‑built agents organized into “companies” (e.g., CEO, Product Manager, CTO). Agents can chat with each other, send “mail”, and participate in multi‑agent chat rooms.
9. Extensible workflows – Built‑in visual Workflow Editor lets you pick a template or design your own graph of steps, gates, and model assignments without touching code.
10. Installation – Zero‑install via npx runfusion.ai, Homebrew formula, or a global npm package. A one‑line installer script also sets up the CLI and dashboard.

Why it matters – Fusion tries to close the gap between high‑level product ideas and actual code by automating the repetitive parts of software development. It gives teams a way to harness multiple LLMs in a coordinated fashion, keep every change isolated in its own Git worktree, and retain human control over critical decisions.

Current status – Early preview, shipping weekly updates. Open‑source under the MIT license, with the core CLI and dashboard available on npm and GitHub.

---

Key take‑aways

• Multi‑agent orchestration for end‑to‑end code generation and review.
• Visual Kanban board + workflow editor for full transparency.
• Works with any LLM provider; can run locally or in the cloud.
• Isolated Git worktrees prevent merge conflicts and enable parallel tasks.
• Extensible via custom agents, workflows, and storage back‑ends.