Tried ChatGPT Apps and Agent Builder.
In ChatGPT Apps, I expected worse, but there’s enough freedom:
- for mcp, both sse and http-streaming are supported
- the entire “app” part is implemented via metadata
- widgets are published through mcp-resources and inserted into chatgpt via an iframe sandbox
- widgets are regular html: you can connect external libraries and so on
- widgets can call tools themselves, change modes, trigger a redirect to an external URL
- three display modes to choose from: inline, fullscreen, picture-in-picture
- tools are regular mcp tools, a widget is attached to each one where the result will be displayed
- development through regular chat is not very convenient due to widget caching, and there are no specialized tools yet
- no app store or publication procedure yet
- the annoying “are you sure you want to call an external tool” window is still there If apps are suggested to the user, it will be interesting.
Agent Builder:
- minimum set of nodes
- no node for arbitrary code or formulas
- the transformation node is quite poor
- instead of access to the call context and data from previous nodes, a “set state” node is used It won’t kill any n8n or zapier because there isn’t a wide range of methods for requesting external data and nodes for data processing. It’s just a visual editor for agent schemas where they forgot to add memory, parallelization, and so on.
UPD: I take back my words about the transformation node. It supports Common Expression Language, which, in theory, should be enough for transformations.