Build and self-host AI agent workflows without the enterprise paywall

What it does

Heym is a visual workflow builder and runtime built from the ground up around LLMs and agents. You drag nodes—agents, vector stores, web scrapers, HTTP calls—onto a canvas, wire them together, and deploy the graph on your own infrastructure. The platform handles multi-agent orchestration, human-in-the-loop pauses, RAG pipelines, and MCP connections natively.

The interesting bit

The README makes a pointed case against “enterprise gatekeeping,” arguing that features like execution history, observability, and cost tracking are core primitives, not upsell bait. Architecturally, it treats AI as the execution model rather than bolting LLM nodes onto classic trigger-action logic. Subagents can also run in parallel behind an orchestrator, which keeps context windows smaller and avoids forcing a single model to juggle unrelated tasks.

Key highlights

• Dual authoring modes: Build workflows by dragging 30+ node types on a Vue Flow canvas, or describe the desired agents and orchestration in natural language and let the system generate the graph.
• Agent depth: Agent nodes support tool calling, persistent memory via per-node knowledge graphs, custom skills backed by generated SKILL.md files, and guardrails for content filtering.
• MCP bidirectional: Agent nodes can act as MCP clients consuming external servers, while the platform itself can expose workflows as an MCP server for tools like Claude or Cursor.
• Production-minded defaults: Built-in LLM trace inspection, token-level cost tracking in USD, one-click evals, and an HITL node for human approval mid-run.
• Auto-healing scrapers: A Playwright Auto Heal feature attempts to fix broken selectors at runtime, which is either clever or a sign of how fragile browser automation remains.

Caveats

• The license is MIT with a Commons Clause rider, making the project source-available but not fully open source for commercial redistribution.
• The competitive comparison table is self-reported and relies on footnotes like “limited” and “partial” for rival platforms; treat it as positioning rather than an objective feature audit.

Verdict

Teams that want n8n-style automation but need first-class agent orchestration, RAG, and cost observability in a self-hosted package should look here. If you need a simple, non-AI trigger-action tool or a fully permissive OSS license, this is not your stack.