An excellent summary of AI in 2025 from Simon Willison.

One point that I’m unsure about is the potential end of MCP: I agree that for coding purposes it doesn’t make sense (if you have a service without a CLI tool, build a great CLI tool, not an MCP server, for it). However, I still wonder if there is use for non-coding purposes (I’ve been experimenting with this for opal3). I don’t trust AI running my browser (as Simon explains), so how else do I interact (as a non-coder) with a closed system? Maybe there should just be a .wellknown link with a skill, but I still need to help the agent authenticate somehow. But will people use “custom connectors” in Claude, and similar features in other agents?