Overview

agentkit is the command-line tool for AgentKit, built around the full lifecycle of an agent: scaffold a project, build and deploy its runtime, invoke the deployed agent, and manage the MCP services, knowledge bases, memory, sandboxes, skills, and evaluation it relies on — all from your terminal.

Quickstart

Install, log in, and deploy your first agent in three steps.

Command reference

Every command with a one-line description and an example.

Command groups

Group	What it manages
`init`	scaffold a new agent project
`config`	manage project lifecycle config and user-wide defaults
`build`	build a runtime image from lifecycle config
`deploy`	deploy an already built runtime from lifecycle config
`launch`	build and deploy a lifecycle project in one command
`status`	show runtime state for the current project
`release`	run the full cloud release pipeline from `.agentkit/agentkit.yaml`
`runtime`	deployed runtimes (list, show, logs, versions, release, update, delete)
`knowledge` (`kb`)	knowledge bases
`memory` (`mem`)	memory collections
`model-gateway` (`mgw`)	unified model gateway, model providers, and consumers
`harness`	Harness defined by a `harness.yaml`
`invoke`	call a deployed runtime
`auth`	SSO login, identity, login profiles, and organization SSO setup
`onboard`	install bundled skills into local coding agents
`eval`	evaluation datasets, evaluators, and experiments
`migrate`	migrate an existing agent application to AgentKit
`mcp service`	create, inspect, and delete MCP services
`sandbox`	run commands and manage sessions in a sandbox
`skill`	skills on the AgentKit skill platform
`destroy`	delete a deployed runtime
`upgrade`	check for and install AgentKit CLI updates
`tree`	print the command tree
`docs`	open or print the AgentKit CLI documentation URL

Conventions

agentkit may also be invoked as ak.
login, logout, and whoami are available as top-level commands and as auth login, auth logout, and auth whoami.
knowledge has aliases kb and knowledgebase, memory has alias mem, and the eval prefix is optional (agentkit dataset … is equivalent to agentkit eval dataset …).
Most read commands accept --json; the region is auto-detected unless you pass -r, --region.
Every product command accepts global --provider volcengine|byteplus to override the cloud provider for that command.
Destructive commands (delete, remove, destroy) prompt for confirmation; pass -y to skip.

Getting Started

AI-Friendly

Command reference

Quickstart

Command reference

Command groups

Conventions

Quickstart

Command reference

​Command groups

​Conventions

Command groups

Conventions