How AgentKit Works
From a verified Kit release to reviewed artifacts — and the boundary between install and execution.
Flow
Seven end-to-end steps
Read in order — each step belongs to a different part of the system.
- 1
Authenticated registry
Provides a verified Kit release — the trusted source for workflow packages.
- 2
AgentKit CLI (ak)
Authenticates, resolves, verifies, installs, updates, diagnoses, removes AgentKit content.
- 3
Runtime adapter
Projects supported content into the coding agent’s structure (Claude Code, Codex, Cursor…).
- 4
Coding agent
Discovers installed Skills, Agents, Hooks, rules, and supporting artifacts.
- 5
Skill / Agent workflow
You invoke a Skill or ask the runtime for an outcome.
- 6
Project artifacts
Code, plans, tests, reports produced by the workflow in your workspace.
- 7
Your review
Decide what to keep, commit, publish, or send. AgentKit does not replace you.
Ownership
Installation and execution stay separate
AgentKit owns the install lifecycle
The CLI manages files it can identify as AgentKit-owned: install, update, remove, and recovery snapshots when available.
The runtime owns the AI session
Model access, tools, permissions, and lifecycle events belong to the coding assistant — not the CLI.
Scope
Project vs user (global)
Scope answers where content is available — not who owns every file in the repo.
Project scope
Available in the current workspace. Use when the Kit should apply to one project.
--global User scope
Runtime user directories — cross-project. Updating one scope does not delete the other.