RondoFlow

RondoFlow is a local-first, MIT-licensed platform for building and running multi-agent AI workflows on a visual drag-and-drop canvas. It orchestrates real Claude Code CLI agents — not just chat completions — so agents can read and write files, run shell commands, and use MCP tools, all governed by per-agent security policies you control. The project began as a fork of Orchestra (also MIT) and has diverged substantially since its initial release.

What It Is

RondoFlow sits in the multi-agent orchestration category, filling the gap between a single Claude Code terminal session and a full production workflow platform. Where Claude Code is one agent taking one turn at a time, RondoFlow lets you wire many agents together on a React Flow canvas, run them as a DAG with parallel branches and conditional edges, and have AI helpers plan, steer, and review each run automatically. It runs entirely on your own machine — your code, prompts, and API keys never leave it except for the model API calls you configure.

How the Canvas and Orchestration Work

The core interaction is a drag-and-drop board where you place node types — Assistants (agents), Skills, Safety Rules, Resources, MCP Connections, Output, Condition (branching), and Sticky Notes — and connect them into a directed acyclic graph. Three AI helpers reason about the run at different points:

• Planner (before the run): reviews the team, models, skills, and wiring, then suggests improvements up front.
• Director (during the run): after each step, decides whether to continue, redirect (retry with sharper instructions), or conclude early.
• Advisor (after the run): scores the finished run and recommends better personas, skills, and step ordering for next time.

A Workflow Generator can also design a 2–5 agent team from a plain-English task description, laying it out as a DAG for review before it hits the canvas.

Bundled Catalog and Building Blocks

RondoFlow ships with 2,438 ready-to-use building blocks: 57 agents, 21 workspaces, 18 facilitators, 2,317 skills, and 25 canvas templates. Pre-built workspace presets include a PR Review Crew (4 agents running correctness and security reviews in parallel), a Deep Research Brief pipeline (adversarial source verification), and an SEO Product Descriptions crew. Skills are reusable instruction sets attachable to any agent; additional skills can be installed from any Git repository. The catalog is community-extensible via GitHub issues.

Architecture and Tech Stack

The server is a Fastify + Socket.IO backend that walks the workflow DAG, runs each agent through the appropriate provider, and gates risky actions behind a three-layer policy model (global, per-agent, per-session — most restrictive wins). The frontend is Next.js with React Flow. PostgreSQL (via Prisma) stores workspaces, audit logs, and analytics. Supported agent providers are Claude Code (local CLI subprocess), OpenAI (GPT/o-series), and Perplexity (Sonar family for web-grounded research); Sakana AI is also callable as a workflow node. The full stack runs in five Docker containers or locally with Node 20+ and Docker for Postgres only.

Update: v1.2.2

The latest release is v1.2.2, published in September 2026. The repository was created in June 2026 and has seen active development, with the last push in September 2026. The project is versioned at 1.0.0 stable per its README badge, with patch releases following. The roadmap and milestones are tracked publicly on GitHub, and contributions are welcome under the MIT license.