Drag-and-drop pipelines for AI agents that read the web

What it does

Open Agent Builder is a browser-based node editor for assembling multi-step AI workflows. You drag nodes—Agent, MCP Tool, If/Else, While Loop, User Approval—onto a canvas, wire them together, and run them with real-time streaming logs. Under the hood, LangGraph handles state management and execution, while Firecrawl supplies the native web-scraping and search primitives.

The interesting bit

The project treats MCP (Model Context Protocol) as first-class plumbing: Anthropic Claude can natively call Firecrawl and custom MCP servers directly from an Agent node, effectively turning the visual graph into an interactive tool-use loop. That makes it less a static DAG and more a live reasoning environment where the LLM decides when to scrape.

Key highlights

• Eight core node types including While Loop and User Approval for human-in-the-loop gates.
• Built on Next.js 16 (canary), Convex for real-time persistence, and Clerk for multi-user auth.
• Sandboxed code execution via E2B for Transform nodes that need custom data manipulation.
• Supports multiple LLM backends—Anthropic Claude, OpenAI GPT-5, and Groq—with per-user API key management in the UI.
• Pre-built templates such as “Simple Web Scraper” and “Competitive Analysis” to bootstrap common pipelines.

Caveats

• The README notes the project is actively under development and some features are still in progress.
• MCP support is currently limited to Anthropic Claude; OpenAI and Groq integrations are marked as “coming soon,” as are Gemini and Open Router.
• Self-hosting requires several external accounts—Convex, Clerk, and Firecrawl—to function.

Verdict

Teams already using Firecrawl who want a shareable, visual interface for complex scraping-and-research pipelines will find this a natural upgrade from scripts. Purely code-first developers who prefer raw LangGraph or plain orchestration YAML will likely find the no-code abstraction and external service dependencies more friction than help.