GitHub

Open source · Local-first

Orchestrate Claude Code agents on a canvas.

Drag agents onto a board and wire them into a workflow. A Planner tunes it before the run, a Director steers each step, and an Advisor reviews the result - all on your own machine.

Install in 3 stepsView on GitHub
  • Open source
  • Runs locally
  • Super easy to use
1Before the run
PlannerBefore

Reviews the agents and how they're wired before you hit Run, then suggests changes up front.

2During the run
DirectorDuring

After each step, decides whether to continue, redirect that step, or conclude the run early.

Task

Your prompt or trigger

Researcher
Instructions
Search the web and gather sources on the topic.
OpusPlan
Output
Data Analysis

Reusable skill

Skill
data-analysis
Skill
Synthesizer
Instructions
Merge the findings into one cited brief.
SonnetEdit
Output
Result

Combined run output

Markdown
3After the run
AdvisorAfter

Scores the finished run and recommends better personas, skills, and step ordering for next time.

  • Plannerbefore
  • Directorduring
  • Advisorafter

Batteries included - 2397 ready-to-use building blocks

57
Agents
11
Workspaces
8
Facilitators
2317
Skills
4
Templates

Built on tools you already trust

  • Claude Code
  • Next.js
  • React Flow
  • Fastify
  • PostgreSQL
  • Prisma
  • Socket.IO
  • TypeScript

Everything it does

One canvas, the whole toolkit

From building and securing workflows to running, scheduling, and reviewing them - here is the full set of what RondoFlow can do.

Build visually

  • Drag-and-drop canvas
  • A dozen+ node types - agents, skills, policies, MCP, resources, notes
  • Conditional branching & parallel DAG runs
  • Ready-made canvas templates
  • Generate a workflow from a task description

Agents & orchestration

  • Multi-agent workflows with parallel branches
  • Planner - analyzes and tunes before the run
  • Director - steers each step mid-run
  • Advisor - reviews the run and suggests fixes
  • Multi-agent discussions with facilitators
  • Claude Code, OpenAI, Perplexity, and Sakana AI

Skills, tools & integrations

  • Reusable skills attached to any agent
  • MCP server connections
  • Web search & deep research
  • HTTP requests & webhooks
  • Send email over SMTP
  • Save to a database, structure & template output
  • Local folders as resources

Guardrails & control

  • Per-agent security policies (most-restrictive-wins)
  • Permission modes - plan, edit, full
  • Tool allow / block lists
  • Approval gates mid-run

Automate & observe

  • Cron-scheduled runs
  • Live run status, streamed
  • Activity feed, audit log & analytics
  • Re-attach to a running workflow

Own your stack

  • Local-first, self-hosted with Docker
  • Workspaces & tabs
  • English, French, German, Slovak, Spanish
  • Dark & light mode
  • MIT-licensed & catalog-extensible

Why a canvas

Claude Code is one agent. RondoFlow runs the orchestra.

Claude Code in the terminal is a single agent taking one turn at a time. RondoFlow keeps that power and adds the layer above it: many agents, wired together, with guardrails and AI that plans and reviews each run.

claude - terminal
$ claude "review this PR"
> reading files...
> one agent, one session
> you drive every turn by hand

done. now do it all again
for the next task...

One agent. One terminal. Manual, every time.

RondoFlow - canvas
Planner
ReviewerResearcherWriter
Advisor

3 agents running in parallel - wired, policed, repeatable

A crew on one canvas. Wired, guarded, and rerunnable.

Many agents, in parallel

Run a whole crew at once - reviewers, researchers, writers - instead of one session, one turn at a time.

Wired into a workflow

Outputs flow agent-to-agent on a visual DAG with branches and conditions, so a run is repeatable - not a chat you'd have to redo.

Per-agent guardrails

Each agent gets its own security policy - allowed tools, paths, and permissions - instead of one blanket trust level for the terminal.

Plan, steer, schedule

Planner, Director, and Advisor plan, steer, and review each run; the Scheduler reruns it on a cron - none of which a bare terminal gives you.

See how RondoFlow compares, point by point

What you can build

Drop in a crew and go

Each of these is a shipped workspace preset - a wired multi-agent pipeline you can add to the canvas and run as-is, then tweak.

PR Review Crew

Point it at a diff or PR — it summarizes the change, runs correctness and security reviews in parallel, then synthesizes one prioritized review report.

4 agentsSee in catalog

Deep Research Brief

Give it a research question — it gathers sources across the web, extracts concrete claims, adversarially verifies them, then writes a cited brief built only from what survived.

4 agentsSee in catalog

SEO Product Descriptions

Feed it product details and target keywords — it rewrites descriptions to be persuasive and keyword-optimized, adds meta + schema, and QA-checks the result.

4 agentsSee in catalog

Install

Up and running in 3 steps

RondoFlow runs locally. You need Node, Docker, and a Claude credential - either an Anthropic API key or a Claude Code OAuth token.

  • Node.js 20+
  • Docker
  • A Claude credential
  1. 01

    Clone the repo

    Grab the source and step inside.

    $ git clone https://github.com/rondoflow/rondoflow.git$ cd rondoflow
  2. 02

    Set it up

    One command installs dependencies, writes your .env, starts Postgres, then migrates and seeds the database.

    $ npm run setup
  3. 03

    Start it

    Launches everything. Open http://localhost:3000 and sign in.

    $ npm run dev

Prefer plain Docker?

Copy the example env, fill in your secrets, and bring the whole stack up in containers instead.

$ cp .env.example .env$ docker compose up

Screenshots

The real canvas

Agents, skills, and guardrails as nodes; live status as a run executes; results streaming back - in a local-first app you own, in whichever theme you like.

localhost:3000
The RondoFlow canvas: agent and skill nodes connected by edges into a workflow, with a run toolbar and a results panel.LightDark

Catalog

Start with what ships in the box

2397 ready-to-use building blocks come bundled - agents, workspaces, skills, facilitators, and canvas templates. Use them as-is, remix them, or build your own.

Pre-built specialists - frontend, data, SEO, research, and more.

Frontend Dev

Engineering

Builds and refines UI with React, TypeScript, and modern CSS.

Backend Dev

Engineering

Designs APIs, data models, and server-side logic.

DevOps Engineer

Engineering

Automates CI/CD, containers, and infrastructure.

Code Reviewer

Engineering

Audits diffs for bugs, quality, and best practices.

QA / Test Engineer

Engineering

Writes tests and hunts for regressions and edge cases.

Data Analyst

Data & Analytics

Explores datasets and turns numbers into clear insights.

Data Scientist

Data & Analytics

Statistical modelling, experiments, and rigorous analysis.

SEO Specialist

SEO & Search

On-page, technical, and content SEO recommendations.

SERP Search Agent

SEO & Search

Pulls search rankings and analyses the results page.

Web Search Agent

SEO & Search

Fast, focused web search and source gathering.

GEO Agent

SEO & Search

Generative Engine Optimization for AI answer engines.

Keyword Analysis Agent

SEO & Search

Researches keywords, intent, and clustering opportunities.

Showing 12 of 57 - browse the full catalog below.

Browse the full catalog

Missing something?

The catalog is open and community-extensible - suggest a new agent, skill, or workspace and we’ll review it.

Suggest a catalog item

Features

Everything you need to wire agents together

A focused toolkit for building, securing, and running multi-agent workflows - without leaving your machine.

Visual canvas

Drag agents, skills, and guardrails onto a board and connect them. The graph you draw is the workflow that runs.

Multi-agent workflows

Branch work across agents in parallel, add conditions, and merge results. RondoFlow runs the DAG and shows each step live.

Reusable skills

Attach instruction sets to any agent. Write your own or pull ready-made ones from the shipped catalog.

Guardrails & policies

Set tool permissions and safety policies per agent. When rules overlap, the most restrictive one always wins.

Scheduled runs

Put any workflow on a cron schedule and let it run on its own - reports, reviews, and research while you are away.

Local-first

Everything runs on your machine against the Claude Code CLI. Your code, prompts, and keys never leave it.

Compare

How RondoFlow stacks up

The short version against Langflow, Flowise, Hiveflow.ai, and Relay.app - see the full 15-point breakdown on the comparison page.

RondoFlowLangflowFlowiseHiveflow.aiRelay.app
Open source
Runs coding agents (your files + shell)
Per-agent security policies
Plan · steer · review the run
See the full comparison

FAQ

Questions, answered

Is RondoFlow free?

Yes. RondoFlow is open source under the MIT license and free to run yourself - you only pay for your own model usage on your Anthropic, OpenAI, or Perplexity account.

Do I need an API key or a Claude subscription?

You need one Claude credential: either an Anthropic API key or a Claude Code OAuth token from `claude setup-token`. That credential is the only thing forwarded to the agents you run.

Does my code leave my machine?

RondoFlow is local-first: it runs on your machine and executes the Claude Code CLI against your own files and tools. Your code, prompts, and keys stay local - only the messages you send to a model reach that model’s provider.

Which operating systems are supported?

Anywhere Node 20+ and Docker run - macOS, Linux, and Windows via WSL.

How is this different from using Claude Code directly?

Claude Code is a single agent in your terminal. RondoFlow orchestrates many of them on a visual canvas - parallel multi-agent workflows, reusable skills, per-agent guardrails, scheduling, and AI that plans, steers, and reviews each run (Planner, Director, Advisor).

Can I use models other than Claude?

Yes. Alongside Claude Code, agents can run on OpenAI and Perplexity - including web search and deep research - and workflows can call Sakana AI mid-run with the Sakana AI node.

Open source

Yours to read, run, and reshape

RondoFlow is MIT-licensed and built in the open. Fork it, file an issue, or send a pull request - the whole platform, including this page, lives in one repository.

Star on GitHubRead the docs
  • MIT license
  • Contributions welcome
  • rondoflow/rondoflow