I made a vim extension where you describe the edit/action you want in natural language, and my ollama model thats trained on books like Practical Vim returns the key sequence and you can press e to execute without leaving vim. So you get automation help but also learn the syntax.
I used mcp-neovim-server to let my LLM control my vim session. that way it can debug and test and poke around. It's crazy good at debugging plugins. It's absurd how little these things need to look-up docs; half these models are just out the box wildly good at vim. "Open these files on these lines" and now there are four splits with me at the relevant line numbers. Awesome. And it'll explain how to do things & test it out & validate! https://github.com/bigcodegen/mcp-neovim-server
That was a little tricky to set-up. I ended up writing nvim-auto-listen, which uses some heuristics to find your project root, and starts a .nvim.socket in that directory. That makes it easy for mcp-neovim-server instances to find. https://github.com/rektide/nvim-auto-listen/
I'm only somewhat getting started, but the workmanship, fit and finish is just outstanding on Codecompanion, for a fantastically well put together in vim agentic experience. Works really well driving a headless opencode mcp. Being able to stay in vim but still get a great opencode powered experience has been mind blowingly sick.
https://github.com/olimorris/codecompanion.nvim
