RondoFlow is an open-source platform that lets developers build teams of Claude Code agents using a visual, drag-and-drop canvas. Rather than wiring prompts together, it orchestrates actual AI agents that can read and write files, run commands, and use MCP tools and skills - all running locally on your machine. The project is hosted on GitHub under rondoflow/rondoflow and is written in TypeScript across a Next.js frontend and a Fastify backend.

How it works under the hood

RondoFlow runs five services when started locally: a Next.js frontend on port 3000, a Fastify backend with Socket.IO on port 3001, a PostgreSQL database on port 5432, a Nextra documentation site on port 3002, and a one-off migration container. Developers can install it either through npm run setup, which handles dependencies, generates the environment file, and starts Postgres, or via docker compose up, which builds and runs all five containers. Docker mode requires Docker Desktop but no Node.js on the host machine.

The core execution engine treats workflows as directed acyclic graphs. A ChainExecutor walks the DAG, dispatching each step to the appropriate agent runner - either the Claude Code CLI as a subprocess, or the OpenAI and Perplexity APIs. Every step streams back live through Socket.IO, showing text output, tool calls, and token usage in the browser.

The three AI assistants that steer every run

What sets RondoFlow apart from standard prompt-chain tools is a layered AI reasoning system that operates at three points during a workflow. A Planner reviews the team composition, models, skills, and execution order before a run begins, suggesting improvements up front. A Director evaluates the output of each step while the workflow is running and decides whether to continue, redirect with sharper instructions, or conclude the chain. It also banks learnings for future runs and supports a tunable criticism level. An Advisor reviews the final result against the original objective and offers one-click fixes.

The platform also supports a Facilitator mode where multiple Assistants brainstorm, review, or debate a topic, with the Facilitator managing turn order and synthesizing a conclusion.

Multi-provider support and the visual canvas

RondoFlow supports three provider tiers. Claude Code is the default and runs the local CLI with full access to tools, MCP connections, and skills, offering Opus, Sonnet, and Haiku model tiers. OpenAI supports GPT and o-series models with optional web search. Perplexity supports the Sonar family for web-grounded research. The Workflow Generator automatically picks a sensible model for each agent, but developers can override this at any time.

The visual canvas is built on React Flow and supports node types for Assistants, Skills, Safety Rules, Resources, MCP Connections, Output, Conditions for branching, and Sticky Notes. Connection types include flow edges for execution order, association edges for configuration, and conditional edges for branching logic. The canvas supports drag-and-drop, undo/redo, keyboard shortcuts, and a command palette. Everything auto-saves. Workspaces are project-based and tied to a folder on the local machine, with export/import for sharing.

Security, roles, and team management

Security operates on three policy layers - global, per-agent, and per-session - where the most restrictive policy always wins. Risky commands require explicit user approval, and budget limits prevent runaway costs. The full security model is documented in SECURITY.md.

RondoFlow is invite-only with no open self-registration. An admin creates accounts through a Users panel and assigns initial roles. Three global roles exist: viewer (read-only, with a locked canvas and hidden palette), editor (can create, edit, delete, and run workflows, including using the Director, Planner, Advisor, and Discussions features), and admin (everything an editor can do plus user management and global settings). Role enforcement happens server-side on both the REST API and the realtime socket layer, and admin actions are written to an audit log.

Additional features for production use

Beyond the core canvas, RondoFlow includes recurring cron schedules, iterative loops that re-run an agent until a goal is met, an in-app Git panel supporting status, branches, commits, and pushes, persistent Memory that carries facts across runs, external folder mounts, and an audit log with an analytics dashboard for monitoring and cost tracking. The Email node can send workflow output via SMTP, configured either through environment variables or at runtime through the Settings panel.

Getting started

On first sign-in, an onboarding wizard walks users through selecting a working directory and a work mode. Quick Start describes a task in plain English and lets RondoFlow generate a 2-5 agent workflow with personas, models, and skills laid out as a DAG. Full Control lets developers assemble agents manually. Built-in templates are available for Code Review, Content Team, Research, and Brainstorm. Skills are reusable instruction sets shipped with the platform and can also be installed from any Git repository. The full documentation is available at docs.rondoflow.app.