The n8n alternative built agent-first, not API-first

What it does

This is a self-hosted, visual workflow builder that treats AI agents as the primary building blocks rather than afterthoughts. You drag nodes onto a canvas to connect LLMs to email, calendars, messaging apps, Android phones, browsers, and payment services, then deploy them as durable background listeners or on-demand webhooks. It ships with 146 nodes across 31 categories and runs either against cloud provider APIs or entirely offline via Ollama and LM Studio.

The interesting bit

The project thinks in org charts, not just data pipelines: an Orchestrator agent can queue tasks for specialist sub-agents and require approval before they run, turning a workflow into a literal team hierarchy. That is a genuinely unusual layer of bureaucracy to bake into an automation tool, and it suggests the authors expect these agents to stick around long enough to need management.

Key highlights

• First-class local LLM support with automatic capability detection for Ollama and LM Studio, so nothing forces you to send data to a cloud API.
• Built-in RAG pipeline (PDF parsing, chunking, embedding) with vector stores including ChromaDB, Qdrant, and Pinecone.
• An Android Agent that pairs with a real phone over QR code and exposes 16 device services—battery, camera, sensors, media playback—to any workflow.
• A unified social node that normalizes incoming messages from WhatsApp, Telegram, Discord, and Twitter/X into one format.
• Memory and context nodes that persist conversation state and agent-explicit facts across trigger firings, inspectable live from the canvas.

Caveats

• The “real Chrome” browser automation harness is explicitly labeled alpha in the README.
• The provider table includes model names such as GPT-5.6 and Claude Opus 5 whose status is unclear from the README alone; they may be placeholders, typos, or future targets.
• The stack requires both a Node.js frontend and a Python backend, so it is not a lightweight single-binary deployment.

Verdict

Anyone running a self-hosted automation stack who is tired of duct-taping LLM calls between HTTP nodes should look here; if you just need a simple IFTTT replacement for lightbulbs, the 146-node kitchen sink will feel like overkill.