I see quite a few requests for MCP support, local model support etc. The next people to ask this will likely be asking for skill support in Hermes, OpenClaw, Pi and other coding harnesses as well.
I know you're a small company with limited time, but MCP doesn't make much sense here (that's a method for inference to request from other hosted models, not endpoints or API access) however agents that are able to work on their own are proving to be very useful. Having Hermes or OpenClaw etc work directly in Magic might not be as easy, but having a skill to test and iterate things directly in a harness and then deploy that into Magic would likely work very well, as it's a massive timesaver having models iterate and only checking the output once a certain bar has been achived.
I would have put this under an existing thread, but it would pay to look ahead. There are too many locally hosted models, and too many ways to host them, for that level of integration to not be a support headache. Hermes skills (or etc) pushes almost all of that back onto the open source network around hermes with only the portion you expose handled on your end.
Watching this space.