Skip to main content

Overview

AgentKit is a capability package manager for AI coding agents. It turns “guess and check” installs into deterministic, reusable recipes.

TL;DR

AgentKit helps agents install capabilities (not just packages) using declarative, multi‑step recipes.

Think: brew/npm for capabilities, not just binaries.

What it is​

  • A CLI for search, inspect, and add
  • A recipe engine for deterministic installs
  • A GitHub‑native registry of capabilities

What it is not​

  • Not an agent runtime
  • Not a chatbot
  • Not an MCP replacement

Mental model​

  • MCP lets agents use tools at runtime
  • AgentKit lets agents install and configure capabilities ahead of time

A capability is a usable workflow, not just a package: it can include installs, configs, downloads, safety checks, and validation steps.

Why it matters​

Multi‑step workflows beat single installs. AgentKit chains steps (download → extract → configure → validate) so agents stop guessing and start executing.

Agents burn tokens when they guess tooling. AgentKit makes installs repeatable, transparent, and safe for automation.

Example​

agentkit add web-automation --dry-run --json

Returns a plan with resolved recipes, steps, and changes.

How it works (flow)​