anikonistack

a declarative and reproducible setup for all of your coding agents!

all of your skills, plugins and MCP servers are declared once; each agent gets a small adapter file that maps them onto that agent’s config format. Add an agent, it gets the whole stack. Add a skill, and every agent you want gets it

How it works

skills.nix ─┐
plugins.nix ├─► module.nix ─► agents/*.nix, one per agent ─► ~/.<agent>/…
mcp.nix    ─┘        (agents arg)      claude-code.nix, codex.nix, …

• skills.nix: attrset name -> path. Own skills from skills/, upstream ones from pinned flake inputs.
• plugins.nix: attrset name -> path in Claude Code plugin layout. Agents with a native plugin mechanism load them as plugins; others get the plugin’s skills/ flattened in.
• mcp.nix: programs.mcp.servers. home-manager translates each server into every agent’s own format.
• module.nix: exposes the three as the agents module argument and imports agents/.
• agents/: one adapter per agent, listed in agents/default.nix.
• flake.nix + flake.lock: every upstream repo pinned to a commit. Same lock, same result, any machine.

Adapters currently in the repo are the agents I use. They are examples of the pattern, not the scope.

Install

Requires nix (the package manager!) with flakes enabled! Nix runs on NixOS, any Linux, macOS, and WSL; NixOS itself is not required!

Since this is also a home-manager module, home-manager is also necessary in order to reproduce this setup!

1. install nix (the package manager!)
2. then nix run home-manager -- init to get a starter flake.nix and home.nix, then continue below

# flake.nix
inputs.anikonistack.url = "github:kidskoding/anikonistack";

# home.nix
imports = [ inputs.anikonistack.homeManagerModules.default ];

nix flake lock --update-input anikonistack
home-manager switch --flake .   # or nixos-rebuild switch

Every agent config under $HOME becomes a read-only symlink into the nix store, rebuilt from the pinned inputs.

Adding an agent

One file. Take the agents argument, feed it into whatever the agent’s home-manager module accepts.

# agents/myagent.nix
{ agents, ... }:
{
  programs.myagent = {
    enable = true;
    skills = agents.skills // agents.pluginSkills [ "caveman" "superpowers" ];
    enableMcpIntegration = true;
  };
}

Then add ./myagent.nix to imports in agents/default.nix.

What agents gives you:

field	what
skills	every skill, name -> path
plugins	every plugin, name -> path
pluginSkills [ names ]	the skills/ of those plugins, merged
skillDirs dir	scan a directory for */SKILL.md
statusline	wrapped hooks/statusline.sh with its runtime deps
inputs	the flake inputs, for adapters that need a pinned repo directly

Adding content

what	where
a skill you wrote	dir with SKILL.md under skills/
an upstream skill	flake input, one line in skills.nix
a plugin	flake input, one line in plugins.nix
an MCP server	one entry in mcp.nix

Secrets never go in the repo. Reference them as ${VAR} in mcp.nix and export the variable before launching.

Machine-specific config

Lives in your own home-manager config, not here: trusted directories, which package provides a binary, per-machine context.

programs.claude-code.package = null;   # binary already installed another way

Updating

nix flake update                             # here, commit flake.lock
nix flake lock --update-input anikonistack   # in your home-manager repo

Layout

anikonistack/
├── flake.nix          pins every upstream repo
├── module.nix         module entry
├── skills.nix         skills
├── plugins.nix        plugins
├── mcp.nix            MCP servers
├── agents/            one adapter per agent
├── skills/            own skills
├── hooks/             statusline.sh
├── claude-md/         own CLAUDE.md sections
└── setup.sh           non-nix fallback: symlinks skills/ and hooks/ into ~/.claude