Introduction
Build AI agent teams visually. Describe what you need, and RondoFlow assembles a team of specialized Claude Code Assistants that work together to get it done.
RondoFlow is a local-first, self-hosted platform with a drag-and-drop Workspace (canvas) where you compose, connect, and run multi-agent workflows. You stay in control: it runs on your machine, against your project folders, and risky actions wait for your approval.
Who it’s for
RondoFlow is for anyone who wants more than a single chat window:
- Developers orchestrating code review, refactoring, and test-writing across a real repository.
- Writers and analysts running content, research, and data tasks as a coordinated team.
- Teams who want repeatable, scheduled, auditable automation rather than one-off prompts.
You don’t need to be an expert to start. Describe a task in plain English and RondoFlow designs the team for you — then you can edit everything on the canvas.
RondoFlow runs entirely on your own machine (PostgreSQL + a Fastify backend + a Next.js UI). Your code, credentials, and run history never leave your infrastructure. See Self-Hosting.
The core idea
Most AI tools give you one assistant in one box. RondoFlow gives you a visual Workspace where you place specialized Assistants (agents) as Cards (nodes), wire them together, and run the whole thing as a workflow. Each Assistant can have its own model, Personality (persona), Skills, Memory, and Connections — and they hand work off to each other to complete tasks no single agent would handle as well.
Beyond Assistants, the canvas also has data and output Cards — route work down different paths with a Condition Card, write a run’s combined output to a file, email it, turn agent prose into a typed dataset, and persist it to the database. See Data Cards.
Headline capabilities
| Capability | What it does | Learn more |
|---|---|---|
| Visual Workspace | Drag Assistant, Skill, Safety Rule, Resource, Connection, Note, and data/output Cards onto the canvas and wire them with flow, association, and conditional edges. | Canvas |
| AI-generated Workflows | Describe a task; RondoFlow designs 2–5 specialized Assistants in a DAG and lets you review before adding them. | Workflows |
| Branching | A Condition Card is a pure router: it sends work down labelled branches by matching the upstream Assistant’s output (contains or regex), with an Else fallback — honored in both parallel and sequential modes. | Data Cards |
| Data & output Cards | Structure agent prose into a typed dataset and Save to DB; write the combined run output to a file (Output); or send it as an email (Email). | Data Cards |
| Orchestration | Planner reviews before running, Director coordinates between steps, Advisor analyzes finished runs. | Orchestration |
| Team Discussions | Multiple Assistants brainstorm or debate a topic while a Facilitator (moderator) manages turns and synthesizes conclusions. | Discussions |
| Multiple providers | Run Assistants on Claude Code, OpenAI, or Perplexity. | Providers |
| Safety Rules | Layered policies (global, per-agent, per-conversation) gate risky commands and cap budgets — most restrictive always wins. | Safety Rules |
| Schedules | Run Assistants or workflows on a cron schedule (e.g. code review every weekday at 9am). | Schedules |
| Loops | Re-run an Assistant iteratively until a goal is met or an iteration cap is reached. | Loops |
| Memory | Give Assistants persistent context that carries across runs. | Memory |
| Skills | Instruction sets that grant specialized abilities; ships with built-ins and supports importing from Git. | Skills |
| Resources | Attach files, folders, and other inputs that Assistants can use. | Resources |
| Git & Connections | Manage branches, staging, and commits inline; attach MCP-based Connections to extend what Assistants can reach. | Git · Connections |
| Users & Roles | Invite-only team workspace with three global roles — viewer (read-only), editor, and admin. | Users & Roles |
| Monitoring & audit | Watch live execution feedback on the canvas, with activity, audit logs, and token-usage tracking. | Monitoring |
The UI ships in English, Slovak, Spanish, French, and German (switchable in Settings) — though code identifiers, model IDs, tool names, and brand names always stay in English.
Multiple AI providers
Every Assistant has a provider. The default claude-code provider starts (spawns) the Claude Code CLI as a subprocess; the openai and perplexity providers talk to their HTTP APIs directly and carry a per-agent model plus tool toggles (web search, deep research).
- Claude Code — full agentic tool use against your local project. Details
- OpenAI — models such as GPT-4.1, GPT-4.1 mini, GPT-4o, and o4-mini, plus deep-research modes. Details
- Perplexity — Sonar models with built-in web search. Details
How it works
Compose on the canvas
Describe a task in plain English (or pick a template) and let RondoFlow generate a team — or drag Cards onto the Workspace yourself: Assistants, Skills, and Safety Rules, plus data and output Cards (Condition, Output, Email, Structure, Save to DB). Connect them with edges to define execution order and routing. Everything auto-saves.
Run as a workflow
Start the workflow from the toolbar. Each Assistant runs in turn (or in parallel branches), passing context downstream. Condition Cards route work down the matching branch, and Structure → Save to DB steps run as real non-agent steps. Director is auto-enabled for workflows with three or more Assistants to coordinate hand-offs.
Monitor and improve
Watch live feedback on the canvas — active Cards glow, completed Cards show a checkmark. Approve any risky action when prompted. When the run completes, Output Cards write the combined result to a file and Email Cards send it (Email is opt-in and off by default). Afterward, Advisor analyzes the results and suggests improvements.
Agents run as real subprocesses against your working directory and can execute commands. Safety Rules and approval prompts exist to keep that under your control — review the Security model before granting broad access.
Terminology
RondoFlow uses friendly names in its UI. You’ll see these throughout the docs, with the internal term in parentheses on first mention:
| You see | Internal term |
|---|---|
| Assistant | Agent |
| Personality | Persona |
| Safety Rule | Policy |
| Connection | MCP Server |
| Facilitator | Moderator |
| Conversation | Session |
| Workspace | Canvas |
| Card | Node |
| Start | Spawn |
See the full Terminology reference.