Your AI's NixOS hallucinations, cured with 1,030 tokens

What it does

mcp-nixos is a Model Context Protocol server that feeds live NixOS data into AI assistants like Claude. It queries public APIs and search indexes — packages, options, Home Manager configs, nix-darwin settings, Nixvim options, FlakeHub flakes, Noogle functions, wiki articles, and package version histories — then surfaces them through two consolidated tools. The pitch is simple: your AI stops inventing package names that sound plausible but don’t exist.

The interesting bit

The author consolidated seventeen tools into two, explicitly to stay under ~1,030 tokens. That’s unusually disciplined context-window budgeting — most MCP servers sprawl. The nix tool uses a single action parameter to switch between search, info, stats, cache checks, and even browsing local flake inputs from the Nix store. One function signature, many surfaces.

Key highlights

• Queries 130K+ NixOS packages, 23K+ options, 5K+ Home Manager options, and smaller sets for nix-darwin, Nixvim, Noogle, FlakeHub, NixOS Wiki, and nix.dev
• nix_versions tool maps package names to historical versions with nixpkgs commit hashes for reproducible builds
• Binary cache status checks with download sizes via cache.nixos.org
• Runs on any OS with Python 3.11+; no Nix installation required for most features
• Supports STDIO, HTTP, and stateless HTTP transports; includes a Pi coding agent extension
• Packaged in nixpkgs unstable with overlay support for declarative installs

Caveats

• Local flake input browsing (flake-inputs action) requires an actual Nix installation; the rest works without it
• The README notes 675 stars but doesn’t cite performance benchmarks or latency figures for the upstream APIs it queries

Verdict

Worth a look if you use Claude, Cursor, or another MCP client and spend time correcting your AI’s NixOS guesses. Skip it if you don’t touch NixOS or already have a reliable search workflow you trust more than an LLM middleman.