I’ve had several requests, so I wanted to post a link here on the forums to the Github repo for the AI “Iggy” project that was presented at ICC 2025. Feel free to download and use it!
This project is not officially supported by IA, however, we are happy to answer questions about this project within this thread.
Does this require the unreleased MCP module, or can some of it run properly without it?
The solution in this repo has been updated to not require the MCP module since it is not available. It uses Web Dev end points to provide the request/response for the LLM.
However, the MCP module/MCP protocol will certainly make this much easier and more seamless. It will also cut down on the amount of setup that is required for extending tools since you have to define them twice, once in web dev and once in the tool request in n8n(in our case). Another benefit of the MCP module will be maintenance, since the LLM will get the updates automatically from the MCP module primitives when it makes the next call.
I will say, converting a Web Dev endpoint to the new MCP module primitives is pretty trivial when that module is available. And then you can take advantage of all the additional benefits the MCP protocol provides.
Is there any video / demo out there?
The demo of this resource was during ICC 2025. If you had an in-person ticket or a live stream pass you can watch that particular session on the app or on the web platform(there should be a link in your registration email). Typically these videos come available to the general public around the first of the new year after ICC.
As far as a more technical deep dive, that is something that we would like to do at some point. No ETA on that right now. The .readme for that repo has a bit more information on getting up and running.
This is an incredible tool and the demo was mind blowing, great work!
For anyone trying this out with local LLMs the horsepower of your model definitely impacts things compared to the demo.
I had good luck out of the box with the Qwen3.6-35B-A3B model, anything smaller choked on the parameter mapping to the New Artifact MCP Tool call, mixing up what order the parameters were passed in, and/or passing in things like "Alarm Status Table" for the ignition_chat_history_id even though they were correctly received by the n8n webhook. The tag interaction tools worked fine even in smaller models.
The data property also leaves a lot to be desired in smaller models. I didn't see anything in the n8n config for generating the correct JSON structure so it looks like that comes from the LLM itself. Gemini as used in the demo seems to pull up correct structures based on the AI search results from Google so if you are using a smaller model you may need to seed the correct configurations elsewhere in the system.