It’s your job to do everything you can to not get locked into a single provider. Token prices are falling and IPOs are nearing, so it’s more important than ever for the labs to try and lock you in to everything around the model. Even with MCP, which is an open standard, it still leads people to lock in. Yeah you’ll add them all to your first agent, but what about the next? And then next? With the best models changing every couple of months, a bundle of MCP connections (and constant re-authentications) is just one more source of friction that hopefully keeps you stuck juuusst long enough until the LLM cycle makes its way around again. It’s so important to use a gateway (both LLM and MCP) to connect all of your agents and ideally an agent orchestrator on top of that to make switching costs as low as possible. And, while it’s pretty common for AI pilled devs to switch models / providers, many other folks don’t know, want, or care. Doing everything at the gateway layer makes it easier to help everyone use AI cost-effectively. Sometimes, it’s just as simple as defaulting people to an open-source model. This is why at @tailscale we’ve taken Aperture from an LLM gateway to MCP gateway to an agent interface that works with any model and can connect to any other machine or agent on your tailnet from both your laptop and phone. I don’t know exactly what the future of agentic AI will look like, but I know it needs the right amount of flexibility to enable new and impactful use cases that will only exist in so far as we can drive costs lower and lower. youtu.be/s-glGHAa1a4?is=50oQ…