ATK: AI Tool Kit for Developers

ATK is a CLI plugin manager for AI-assisted development.

Install MCP servers and local AI services with one command. Plug them into every coding agent you use (Claude Code, Codex, Gemini CLI, Augment Code, OpenCode) at the same time. Keep your entire setup git-backed, reproducible, and upgradeable.

Install. Plug. Done. atk add github → atk plug github --claude --codex --auggie

Quick start

Requires Python 3.11+ and uv (brew install uv on macOS).

# Install ATK
uv tool install atk-cli        # or: pip install atk-cli

# Initialize ATK Home (~/.atk, a git repo)
atk init

# Browse the registry, install a plugin
atk search
atk add fetch

# Plug it into your coding agents: MCP registration + skill injection in one go
atk plug fetch --claude        # add --codex --gemini --auggie --opencode as needed

# See what's installed and running
atk status

Your entire setup lives in ~/.atk/, a git repository. Push it. Clone it on another machine. Run atk install --all to reinstall every plugin from the manifest. Everything comes back exactly as you left it.

The problem

If you use coding agents seriously, your setup probably looks like this:

• MCP servers installed from random Git repos, each with its own README to follow
• Local services started with long-forgotten docker run commands
• Agent configs hand-edited in JSON files scattered across your machine
• The same MCP configured differently in Claude, Codex, and Augment because you did it three times manually
• Secrets in .env files with no connection to anything

It works. Until you switch machines, break something, want to roll back, or come back after two months and have no idea what's running or how it got there.

ATK exists because this setup is real, fragile, and universal.

What ATK does

Discover and install AI tools from a curated registry

atk search                  # browse vetted plugins
atk add github              # install in one command, prompts for config
atk status                  # see what's running, ports, env status

Plug tools into all your coding agents at once

atk plug github --claude --codex --gemini --auggie --opencode

One command. ATK registers the MCP server with each agent's native mechanism and injects the plugin's skill. No manual JSON editing across multiple apps. Plugins that ship only a SKILL.md (no MCP server) work too, so instruction-only plugins like coding personas are first-class.

Teach your agents how to use the tools

When a plugin ships a SKILL.md (usage instructions for AI agents), atk plug injects it into each agent's context automatically, in each agent's native format (see Supported agents).

Your agent doesn't just have access to the tool. It knows how and when to use it.

Manage the full lifecycle of everything

atk start openmemory        # start a service
atk stop langfuse           # stop it
atk logs openmemory         # tail logs
atk upgrade --all           # pull latest for all plugins
atk remove github           # stop + uninstall + delete

Every tool, whether a Docker service, an MCP server, or a CLI binary, gets the same uniform interface.

Supported agents

atk plug drives all of this: pass a flag for each agent you want to target. atk unplug reverses it.

Registry

The registry currently ships 20 plugins:

Registry plugins are reviewed, schema-validated, versioned, and pinned; registry review is what earns the green verified maturity you see in atk status. Git and local plugins declare their own maturity, and atk add asks for confirmation before installing anything below verified. Search by keyword: atk search memory, atk search git.

Spotlight: claude-dashboard

A live, per-chat executive dashboard for Claude Code sessions:

• A localhost server renders a visual dashboard.html for the active chat and refreshes it after every turn: glance header, to-dos, heads-up items, session journey
• A top-level landing page covers every project and chat with history, so you can find and review any session at a glance
• Runs entirely on your machine; uses your existing Claude subscription, no API key needed. Stop it any time with atk stop claude-dashboard

atk add claude-dashboard

Spotlight: openmemory

Persistent memory for your agents, self-hosted:

• Agents store and recall memories across sessions via MCP, with semantic search
• Local SQLite plus Ollama embeddings; all data stays on your machine, no cloud API
• Ships a SKILL.md so plugged agents know when to store and when to recall

atk add openmemory

Command reference

Three ways to add plugins

1. Official ATK Registry (vetted plugins)

atk add openmemory
atk add langfuse

The "known good" layer: reviewed, versioned, and pinned (see Registry).

2. Git repository plugins (distribution channel)

Any Git repository can become an ATK plugin. Add a .atk/plugin.yaml to your repo and users can install it with one line:

atk add github.com/your-org/your-tool

ATK fetches just the .atk/ directory's contents, validates the plugin, pins it to a commit hash, and manages its lifecycle like any other plugin. This turns ATK into a distribution channel for AI tooling, without a centralized gatekeeper.

3. Local plugins (personal or internal tooling)

atk add ./my-plugin

atk add <path> takes any directory that contains a plugin.yaml. The plugin lives in ~/.atk, fully versioned, same schema. Ideal for personal scripts, internal tools, or plugins in development.

Reproducibility

ATK environments are fully reproducible:

• Plugins are validated against a versioned schema
• Plugin versions are pinned to exact commit hashes in the manifest
• Secrets live in isolated, gitignored .env files
• Every change to ATK Home (add, remove, upgrade) is a git commit; rollback is git revert
• Additive schema changes are backward-compatible

Clone the repo on a new machine. Run atk install --all. You get the same toolchain.

Design principles

Who ATK is for

Developers who:

• rely on coding agents (Claude Code, Codex, Augment Code, etc.)
• don't want vendor lock-in, or already work with multiple agents
• use MCP servers (local and remote)
• run local services like memory, observability, or vector stores
• care about owning their data and controlling their setup

ATK is a git-backed, CLI-first plugin manager for local, long-lived AI tooling, designed to be driven by humans and coding agents.

ATK is not an environment manager (Nix, Conda, Devbox), infrastructure-as-code (Terraform, Ansible), or a production deployment system. It manages one global, machine-level setup, not per-project config.

If you're configuring servers, ATK is the wrong tool. If you're keeping your AI dev setup sane, it's the right one.

For MCP authors: ATK as your distribution layer

If you're building an MCP server, ATK is the easiest way to get it into your users' agents.

One-command install from your repo

atk add github.com/you/your-mcp-server

Add a .atk/plugin.yaml to your repo (see Git repository plugins) and your users get a working, managed, upgradeable installation from one command.

Automatic agent wiring, all agents at once

atk plug your-mcp-server --claude --codex --gemini --auggie --opencode

ATK handles the agent-specific plumbing: native CLI commands for Claude, Codex, and Gemini, JSON config writing for OpenCode, everything. Your users don't need to know which config file to edit or which flags to pass.

You control how agents use your tool

Ship a SKILL.md alongside your plugin. When users run atk plug, ATK injects it into each agent's context automatically (see Supported agents). The agent doesn't just have access, it has instructions. You decide what the agent knows about your tool, how it should use it, and what it should avoid.

How to add ATK support to your repo

You don't need to write the plugin files by hand. ATK ships a dedicated skill file that tells your coding agent exactly what to build: the schema, lifecycle scripts, SKILL.md conventions, testing protocol, and all three distribution patterns.

ATK Plugin Creation Skill →

Feed it to your agent and ask:

"Create an ATK plugin for [your tool name]. Follow the skill file."

The agent will produce a complete plugin.yaml, install and lifecycle scripts, SKILL.md, and README.md, ready to ship.

Once the files are ready, test locally then share via your existing repo:

atk add ./.atk                       # install locally to test
atk add github.com/you/your-repo     # users install from your git URL

Getting listed in the registry

The ATK registry is the curated list of plugins available via atk search. Submit a PR to add your plugin. Registry plugins are reviewed and must meet the schema requirements; verified plugins get a green verified maturity in atk status.

Status

ATK is under active development. Expect fast iteration and opinionated choices.

If this problem resonates with you, try it, and break it.

If ATK saves you time, a ⭐ on GitHub goes a long way.