AgentForge

Workflow automation tools make you build the workflow. You pick nodes, wire them together, and encode every decision as a rule up front. AgentForge does neither. You type what you want — “when my form webhook fires, summarise the submission, decide whether it’s urgent, post urgent ones to Discord, and log every one to my Google Sheet” — and it builds a real workflow on a canvas. Real nodes, real connections, saved to your account, editable field by field. Not a mockup and not a chat transcript. Then it runs, and this is the part that is not a flowchart: its agent nodes call other nodes as tools and decide what to do at runtime. Nobody wrote a keyword rule for “urgent”. The agent reads the message, judges it, and the workflow takes that branch. Send a polite feature request instead and it takes the other one. While it runs, per-node status and the agent’s own reasoning stream live to the canvas over SSE — including for a run started by an external webhook while you are just watching. It is deployed, public, and runs on free tiers: one Cloud Run container, Neon Postgres, Gemini.