mario-andreschak/FLUJO

FLUJO – Visual Builder for Private AI Agents

What it is – FLUJO is an open‑source, locally‑hosted platform that lets you connect large‑language‑model providers (OpenAI, Anthropic, Gemini, Ollama, etc.) and external “MCP” services, then compose them into chat‑based agents using either a simple step‑by‑step recipe UI or a drag‑and‑drop flow diagram. All data and API keys stay on your machine (or on a private network) unless you deliberately expose the service.

Key capabilities

• Local‑first security – encrypted storage of secrets, optional password protection, and no secrets ever reach the browser.
• Multi‑model support – configure any OpenAI‑compatible provider, Claude via its SDK, Google Gemini, X.ai, OpenRouter, or local models through Ollama.
• MCP integration – discover, install, and call tools, resources, and prompts from Model‑Context‑Protocol (MCP) servers. FLUJO can also act as an MCP proxy so other MCP‑aware apps (Claude Desktop, Cursor, etc.) reuse the same server configuration.
• Agent builder –
  • Simple mode: write a linear recipe in plain language.
  • Expert visual mode: connect Start, AI, app, sub‑flow, and Finish nodes on a canvas, with branching, loops, and hand‑off tools.
  • Subflows: reuse a flow as a function with its own isolated state.
• Live debugging – step through a run node‑by‑node, set breakpoints, and inspect state before/after each step.
• Automation triggers – schedule cron‑style runs, webhook calls, file‑watch, URL‑watch, or periodic MCP‑tool polling, with full run history.
• Conversation UI – chat with any agent, attach files/audio, edit messages, and view token usage per step.
• OpenAI‑compatible endpoint – expose any flow as http://localhost:4200/v1 so existing code that talks to OpenAI can call flow‑<flow‑name>.
• Built‑in API docs – a searchable /docs page lists every REST endpoint (chat, flows, MCP, triggers, etc.).

How to get it running

1. One‑click installers – Windows users can download flujo‑setup.exe or run the PowerShell one‑liner (irm … | iex). Linux/macOS users can run the curl‑pipe‑bash script.
2. Docker – docker compose up --build builds the image from source and starts the service on port 4200 (plus a sandbox port 4201 for MCP apps).
3. Manual – clone the repo, npm install (or yarn), then npm run dev for development or npm run build && npm start for production.
4. npx shortcut – npx flujo-ai downloads a pre‑built binary and launches it without a clone.

Typical use cases

• Personal AI assistant – combine a LLM with a calendar‑tool MCP server to schedule meetings automatically.
• Customer‑support bot – trigger a flow via a webhook when a new ticket arrives, let the model call a ticket‑system MCP tool, and reply.
• Data‑pipeline orchestrator – use subflows and loops to fetch data from APIs, transform it with a model, and store results in a database via an MCP resource.
• Team collaboration – put several agents in a shared “meeting” conversation so they can brainstorm or divide work.
• Prototype integration – expose a flow as an OpenAI‑compatible endpoint and call it from existing codebases (e.g., a VS Code extension or a web app).

Limitations / things to watch

• No built‑in authentication; the server is bound to localhost by default. Exposing it publicly requires an external reverse‑proxy with auth.
• The OpenAI‑compatible endpoint only supports flow calls, not a full Ollama‑style model server.
• MCP servers run arbitrary code; you should only install servers you trust, and keep the sandbox sandboxed (port 4201).
• While the UI is visual, complex flows can become hard to maintain without good naming and documentation.

Where to learn more

• Official site & product video: https://flujo.com.co/
• Quick‑install section of the README (one‑click installers, Docker, npx).
• Built‑in /docs page once the app is running.
• Community Discord and GitHub Issues for support.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.