Heym

Product Introduction

1. Definition: Heym is a source-available, self-hosted, AI-native workflow automation platform. Technically classified as an LLM Orchestration and Low-Code/No-Code (LCNC) automation environment, it provides a visual canvas for building complex AI pipelines. Unlike traditional automation tools that retrofitted AI as a plugin, Heym is built on an AI-first execution model, utilizing a directed acyclic graph (DAG) engine to manage multi-agent systems, Retrieval-Augmented Generation (RAG), and Model Context Protocol (MCP) integrations.

2. Core Value Proposition: Heym exists to democratize the creation of production-grade AI automations by removing the coding barrier while maintaining enterprise-level control. It focuses on solving the "AI gluing" problem—the friction of connecting LLMs, vector databases, and external APIs. Its primary keywords include AI workflow automation, multi-agent orchestration, self-hosted AI agents, RAG pipeline builder, and no-code LLM orchestration. It enables users to deploy intelligent agents on their own infrastructure, ensuring data sovereignty and zero vendor lock-in.

Main Features

1. Visual Workflow Canvas with AI Assistant: Heym features a drag-and-drop editor powered by Vue Flow, allowing users to wire together 38 different node types. The platform includes a native "AI Builder" assistant that uses natural language or voice-to-text processing to generate workflows automatically. When a user describes a process, the assistant streams the response and applies the generated JSON configuration directly to the canvas, creating nodes, edges, and logic branches in real-time.

2. Multi-Agent Orchestration & Portable Skills: The platform supports hierarchical agent structures with up to five levels of nesting. A primary orchestrator agent can delegate specialized tasks to sub-agents or sub-workflows. Each agent node is equipped with Python tool-calling capabilities, automatic context compression (triggering at 80% of the model's context window), and a "Skills System." This system allows users to drag and drop portable capability bundles (consisting of SKILL.md instructions and Python scripts) to extend agent functionality without manual configuration.

3. Integrated RAG Pipeline & Vector Store Management: Heym includes native management for Qdrant vector stores. Users can upload various document formats—PDF, Markdown, CSV, and JSON—directly through the dashboard. The RAG nodes support semantic search with metadata filtering and optional Cohere reranking for high-precision retrieval. This integration eliminates the need for external vector database middleware, allowing LLMs to access private knowledge bases within the same runtime.

4. Bidirectional MCP Support (Model Context Protocol): Heym acts as both an MCP client and an MCP server. As a client, Agent nodes can connect to external tool servers to gain specialized capabilities. As a server, any workflow created in Heym can be exposed as an MCP tool for external clients like Claude Desktop or Cursor. This enables a "headless" automation strategy where AI editors can trigger complex, multi-step Heym workflows via a standardized protocol.

5. Human-in-the-Loop (HITL) & Observability: To ensure reliability in production, Heym incorporates HITL approval checkpoints. These nodes pause execution and generate public, shareable review links where human operators can accept, edit, or reject AI outputs. Complementing this is a comprehensive LLM Trace viewer that captures every prompt, response, tool call, and millisecond of latency, providing the observability required for debugging and optimizing agent behavior.

Problems Solved

1. Pain Point: Brittle Browser Automations & DOM Changes: Traditional RPA (Robotic Process Automation) often fails when website UIs change. Heym solves this with "Auto-Heal" browser automation. Using Playwright, if a CSS selector fails at runtime, the AI agent uses vision and reasoning to identify the new element and retry the action automatically.

2. Pain Point: Data Privacy and Compliance Risks: Many enterprises cannot use SaaS automation tools like Zapier or Make due to PII (Personally Identifiable Information) regulations. Heym's self-hosted architecture (deployable via Docker Compose or Kubernetes) ensures that all data, LLM traces, and credentials remain within the user's private infrastructure.

3. Target Audience:

• AI Engineers & Developers: Looking for a faster way to prototype and deploy multi-agent systems without writing boilerplate Python code for every integration.
• DevOps Teams: Needing to automate infrastructure tasks using LLMs while maintaining security through self-hosting.
• Customer Support Managers: Aiming to build intelligent triage systems that use RAG to answer tickets based on internal documentation.
• Productivity Power Users: Who require complex logic (loops, conditions, parallel execution) that exceeds the capabilities of simple chat interfaces.

4. Use Cases:

• Automated Support Triage: Monitoring IMAP inboxes, using RAG to find answers, and routing urgent issues to Slack while requiring HITL approval for refunds or account changes.
• Content Generation Engines: Multi-agent workflows where one agent researches via the Google Search/Crawler nodes, another drafts content, and a third applies guardrails for brand safety.
• Real-time Data Processing: Using WebSocket and RabbitMQ triggers to process incoming data streams through LLMs for sentiment analysis or entity extraction.

Unique Advantages

1. Differentiation from Legacy Automation: While tools like n8n or Zapier have added AI nodes, Heym is "AI-native." This means features like parallel DAG execution (running independent nodes concurrently), context window management, and agent reasoning loops are core primitives rather than superficial additions. Heym treats an "Agent" as a stateful entity rather than just a linear prompt step.

2. Key Innovation: The Portal Chat UI: Heym can transform any workflow into a public-facing chat interface at a custom URL. This "Portal" mode supports file uploads, image rendering, and streaming responses, essentially allowing users to ship a custom AI application (SaaS-lite) directly from an automation workflow without frontend development.

3. No Vendor Lock-in: By supporting local LLM providers like Ollama and vLLM alongside OpenAI and Anthropic, Heym ensures that users can switch models or providers by simply changing a credential, protecting the automation stack from pricing changes or API deprecations.

Frequently Asked Questions (FAQ)

1. Is Heym a free alternative to Zapier or n8n for AI? Yes, Heym is source-available and free to self-host under the Commons Clause + MIT license. It provides many features—such as multi-agent orchestration and built-in RAG—that are often gated behind expensive enterprise tiers in other platforms. It is specifically optimized for AI-heavy workloads that require more flexibility than traditional automation tools offer.

2. How does Heym handle long-running AI agent conversations? Heym utilizes automatic context compression. When an agent's conversation history approaches the model's context limit (typically around 80%), the engine summarizes the middle portion of the conversation while preserving the system instructions and the most recent turns. This prevents "context overflow" errors and allows for stable, long-term interactions.

3. Can I run Heym entirely offline or in an air-gapped environment? Yes. Because Heym is self-hostable via Docker and supports Ollama for local LLM inference, the entire stack can operate without an internet connection. This is ideal for organizations handling sensitive proprietary data that must never leave their local area network.

4. What is the difference between an LLM node and an Agent node in Heym? An LLM node is a single-turn execution point for text or vision generation. An Agent node is an autonomous loop that can use "tools" (Python scripts, MCP servers, or other workflows) to complete a goal. Agents can think, act, and observe results, making them suitable for complex tasks that require multi-step reasoning rather than just a single prompt.