claudecode.nvim

You will like the ergonomics if the shape fits. The agent runs where it always ran, and this gives it eyes: your current selection, the file you are on, the entries you picked in a tree all become things the session can see without you pasting anything. The deciding daily trait is that transfer, because most of the friction in terminal agents is describing where you are.

Pick it if your session is already open in a split. Pick avante.nvim when you want suggestions and edits arriving in the buffer itself.

The dependency here is on an interface nobody agreed to keep. This works by reimplementing what a vendor's editor extensions do internally, which means every capability rests on behaviour that was never a promise, and the vendor is under no obligation to notice this project when it changes. The failure mode is not a bug report, it is a plugin that silently stops connecting after an unrelated update.

What it does right: the protocol is written down in the repository for anyone else to build on, which turns private reverse engineering into shared documentation.

1. The transport is a socket carrying a standard remote-procedure format with full request and response handling, implemented in the editor's own runtime with almost no dependencies. 2. The role assignment is the interesting part: the editor exposes tools and the agent calls back into them, which reverses the usual arrangement and makes the human's cursor a first-class context source. 3. No edits are applied by the plugin, so verification remains wherever it was before.

No benchmark applies to a transport. The observation: pushed context is cheaper and more accurate than retrieved context, and almost nobody builds it this way.

The ownership is worth reading carefully. This sits under an infrastructure company that sells something else entirely, so the plugin is a goodwill exercise aimed at exactly the engineers that company wants attention from. That is a legitimate and cheap channel, and it also means nobody's revenue depends on this shipping, so the risk is not insolvency but indifference.

There is no acquirer, because it is already inside one, and no pivot, because there is nothing to pivot. Position: free distribution from a funded parent, maintained as long as it flatters the parent, which is a shorter horizon than a licence.

The demo takes thirty seconds and proves very little. Sixty seats costs nothing here because this is a bridge, and the spend sits on the underlying agent plan that finance already reconciles. There is no console, no policy surface, no telemetry to disable and no retention question, because the plugin holds nothing. It also contributes nothing to review, nothing to our pipelines and nothing to any control we would be asked about.

Onboarding is one line in a config for the handful of engineers who use this editor. Not yet, as anything other than personal tooling.

I can read every line of this in an evening. It is Lua on the editor's built-in event loop with a single external plugin, so there is no compiled blob, no background service and nothing installed outside my plugin directory. Permissive licence, so the fork is mine forever, and the written protocol means I could point it at something else entirely.

The limitation is at the top rather than the bottom: I cannot bring my own model here, because the whole design targets one vendor's agent. My hardware is irrelevant to it. That is a deliberate scope, and it is still a leash.