Feature: Read flake inputs from local nix store

[dhnaranjo]

Whattup friend.

While working on my configuration, I've noticed that the agent frequently searches GitHub or other repository locations for information about flake inputs (like home-manager, flake-parts, etc.). However, all of these inputs are already present in the local Nix store after they've been added to my flake.

Proposed Enhancement:

Add an MCP tool that can access and query these flake inputs directly from the filesystem/Nix store. This would provide several benefits:

1. Faster responses - No need to download or fetch from web APIs
2. More accurate results - Reading directly from the exact versions pinned in the flake lock file
3. Fewer external dependencies - Reduced reliance on search APIs and network requests
4. Better offline support - Works without internet connectivity once inputs are fetched

The tool could leverage the fact that flake inputs are stored in /nix/store and are accessible locally, providing immediate access to the actual source code and documentation of the dependencies already in use.

Is this something that could be added to the mcp-nixos server?

[jamesbrink]

I think this is a good idea, I need to think about how to best approach this. I don't want the MCP server's context to get too big (ever since claude code added /context command I have been unloading MCP servers 😂 ).

I went into this project with the idea that Nix/Nixos is not even a requirement to use the MCP, but it seems like stuff like this can be added and gracefully handled. Your approach makes a lot of sense, as to do the benifits.

[dhnaranjo]

I've been running with this for a few days, I added an output to my flake mapping the input store paths, taught it how to nix eval .#inputPaths.{{NAME}} --raw and added a filesystem mcp restricted to the store. It definitely does use it, but it also searches for non-inputs in the store, or it'll run find on the whole store, dumb stuff like that. I find that generally if you give the thing multiple similar ways to do a category of thing it's gonna misapply them.

So I dunno... it might be better to just let it find them on GitHub.