mario-andreschak/FLUJO

FLUJO – Visual, local‑first AI agent builder

What it is – FLUJO is an open‑source desktop‑app that lets you wire together large‑language‑model calls, external tools, and custom scripts into agents you can chat with, schedule, or trigger from other software. It stores everything (models, flows, chats) on your computer, only sending the minimal data required to the chosen AI provider or connected service.

Key ideas

• AI Setup – Connect to major providers (OpenAI, Anthropic, Gemini, Azure, Ollama, etc.) via a guided UI. You can save multiple connections, organize them in folders, and define fallback policies.
• Connected Apps (MCP) – Through the Model Context Protocol you attach files, browsers, APIs, GitHub repos, or any locally‑run service as tools that the agent can call. Tools can return interactive UI elements inside the chat.
• Agent Builder – Two modes:
  • Easy: write a plain‑language recipe step‑by‑step.
  • Expert: drag‑and‑drop graph of nodes (Start, AI, Tool, Subflow, Finish). Supports branching, loops, and reusable sub‑flows.
• Talk & Debugging – Run an agent in a chat window, watch each node execute, set breakpoints, and approve tool calls before they run.
• Automations – Schedule agents, fire them on webhooks, file changes, tool‑watch events, website changes, or when another flow finishes. The FLUJO server must stay running.
• Experimental features – Meetings (multi‑agent conversations), Personas (long‑term helpers with memory), and Waves (read‑only overview of flows and timing).
• External integration – Exposes an OpenAI‑compatible HTTP endpoint (/v1) and MCP endpoints so other programs (Cline, Cursor, custom code) can invoke flows.

How you use it

1. Install via the provided Windows installer, PowerShell script, or the Unix install.sh (or run npx flujo-ai@latest).
2. Open http://localhost:4200, add a model connection, and test it.
3. Add tools under Connected Apps (e.g., a web‑scraper, a GitHub‑hosted command).
4. Build an agent – start simple, then optionally switch to the visual graph.
5. Chat with the agent, debug, or set up an automation trigger.

Privacy & security – All workspace data lives locally; you can encrypt it with a password. Only the provider APIs and connected tools receive the request payloads. Network exposure is disabled by default; you must explicitly enable LAN/public access and add authentication.

Who it’s for – Developers, power‑users, or teams that want a self‑hosted, visual way to compose LLM‑driven workflows without writing code, while keeping data under their control.

Status – Version 3.46.3 (MIT‑licensed). Core features (model connections, visual agents, debugging, automations) are stable; meetings, personas, and waves are marked experimental.