What it does Open Agent Builder is a browser-based canvas for stitching together AI agent workflows. You drag nodes—Agent, If/Else, While Loop, User Approval—connect them, and hit Run. Under the hood it uses LangGraph for state management and Firecrawl for scraping, with real-time streaming as the workflow executes.

The interesting bit The project treats MCP (Model Context Protocol) as a first-class citizen: Anthropic Claude gets native MCP tool calling, so an agent node can autonomously decide when to scrape, search, or crawl. Other providers are “coming soon,” which currently makes Claude the only fully supported option for tool-heavy flows.

Key highlights

• Eight node types including human-in-the-loop approval gates and sandboxed code transforms via E2B
• Convex handles real-time persistence; Clerk handles auth; React Flow handles the canvas
• Users can bring their own LLM keys (Anthropic, OpenAI, Groq) and Firecrawl keys through the UI
• Pre-built templates: simple scraper, multi-page research, competitive analysis, price monitoring
• Deployable to Vercel with a standard Next.js build

Caveats

• MCP support is currently limited to Anthropic Claude; OpenAI and Groq are listed as “coming soon”
• The README notes the project is “actively under development” with some features still in progress
• Setup requires accounts and keys for Convex, Clerk, and Firecrawl before you can run anything locally

Verdict Worth a spin if you want a self-hosted, visual alternative to no-code automation tools and your workflows revolve around web data. Skip it if you need mature multi-provider MCP support or a production-grade SLA today.