zhub

Zerone Agent Hub CLI — manage agents, skills, tools, MCPs, and providers from the terminal. Designed for AI agent hosts (Claude Code, Cursor, Cline, etc.) to invoke via shell.

Install

npm install -g @zerone-agent/zhub

Requires Node.js ≥ 18 or Bun ≥ 1.0.

Login

# 1. Create a CLI token in control-panel web UI (User Menu → CLI Tokens)
# 2. Login (non-interactive, all flags required):
zhub login --url https://console.currantmind.cn --token cli_xxxxxxxxxxxx

Command Overview

Global

zhub whoami                    # show current profile + user
zhub --help                    # full help

Agent (14 commands)

# CRUD
zhub agent list [--enabled] [--disabled] [--output json|table|yaml]
zhub agent get <name> [--output yaml]
zhub agent create --file <agent.yaml> [--enabled|--disabled] [--default|--no-default] [--output yaml]
zhub agent update <name> --file <agent.yaml> [--output yaml]
zhub agent delete <name>

# Relations
zhub agent set-subagents <name> <sub1> [sub2 ...]
zhub agent set-tools <name> <tool1> [tool2 ...]
zhub agent set-mcps <name> <mcp1> [mcp2 ...]
zhub agent set-skills <name> <skill1> [skill2 ...]

# Deploy lifecycle
zhub agent deploy <name> [--force]      # deploy to runtime
zhub agent undeploy <name> [--purge]    # archive or permanently delete
zhub agent start <name>                 # restart stopped agent
zhub agent stop <name>                  # stop running agent
zhub agent status <name>                # check deployment status

Provider (6 commands)

zhub provider list [--output json|table|yaml]
zhub provider get <id> [--output yaml]
zhub provider create --file <provider.yaml> | --json '<json>'
zhub provider update <id> --file <provider.yaml> | --json '<json>'
zhub provider delete <id>
zhub provider probe <id>                         # test stored provider
zhub provider probe --base-url X --api-key Y --protocol Z  # test custom config

Tool (5 commands)

zhub tool list [--output json|table|yaml]
zhub tool get <name> [--output yaml]
zhub tool create --file <tool.yaml> | --json '<json>' [--output json|table|yaml]
zhub tool update <name> --file <tool.yaml> | --json '<json>' [--output json|table|yaml]
zhub tool delete <name>

MCP (5 commands)

zhub mcp list [--output json|table|yaml]
zhub mcp get <name> [--output yaml]
zhub mcp create --file <mcp.yaml> | --json '<json>' [--output json|table|yaml]
zhub mcp update <name> --file <mcp.yaml> | --json '<json>' [--output json|table|yaml]
zhub mcp delete <name>

Skill (6 commands)

zhub skill list [--type expert|community] [--output json|table|yaml]
zhub skill get <name> [--output json|table|yaml]
zhub skill create --from-dir <dir> [--name N] [--title T] [--title-en T] [--description D] [--description-en D] [--type T] [--output json|table|yaml]
zhub skill update <name> --from-dir <dir> [--title T] [--title-en T] [--description D] [--description-en D] [--type T] [--output json|table|yaml]
zhub skill delete <name>
zhub skill download <name> [--output json|table|yaml]

Output Formats

All list/get commands support --output:

| Format | Use case | |--------|----------| | table (default) | Human reading in terminal | | json | Agent parsing (always wrapped in { data, meta }) | | yaml | Single-object details, copy-paste friendly |

Agent YAML Format

Agent create and update accept YAML through --file. Agent update does not accept --json; if its YAML provides id, that value must exactly match the positional <name>. Agent create always requires YAML id.

The preferred native shape keeps agent configuration under config:

id: research-agent
enabled: true
isDefault: false
config:
  title:
    zh: 研究员
    en: Research Agent
  description:
    zh: 负责深度研究与总结
    en: Performs deep research and synthesis
  modelId: claude-sonnet-4-5
  systemPrompt: |
    你是一个研究员。
  maxTurns: 20
  permissionMode: auto

For compatibility, the equivalent flat shape is also accepted. model is normalized to config.modelId, while the other configuration fields are moved under config:

id: research-agent
enabled: true
isDefault: false
title:
  zh: 研究员
  en: Research Agent
description:
  zh: 负责深度研究与总结
  en: Performs deep research and synthesis
model: claude-sonnet-4-5
systemPrompt: |
  你是一个研究员。
maxTurns: 20
permissionMode: auto

Do not define the same normalized field in both places (for example, top-level model together with config.modelId, or top-level title together with config.title); conflicting definitions are rejected. title and description must be language-to-string maps.

On create, YAML enabled and isDefault values are used unless overridden by CLI state flags. --enabled and --disabled are an explicit conflict pair and must not be supplied together. --default and --no-default use Clipanion's native last-argument-wins behavior, so the rightmost occurrence determines the value. If omitted everywhere, create defaults to enabled and not default. Update reads state only from YAML and leaves omitted state fields unchanged.

See skills/zhub-management/references/agent.yaml for a full example.

Skill Metadata Flags

Both skill create and skill update accept four localized metadata flags:

• --title: display title
• --title-en: English display title
• --description: Chinese description
• --description-en: English description

skill create falls back to the skill name for --title and to SKILL.md frontmatter for --description; the English fields are optional.

Use with Claude Code / Cursor

The zhub-management SKILL lives in this repository at skills/zhub-management/. Symlink or copy it to your agent host's skills directory:

# From the repository root
ln -s $(pwd)/skills/zhub-management/ ~/.claude/skills/zhub-management

# Or copy
cp -r skills/zhub-management/ ~/.claude/skills/

For Cursor, use ~/.cursor/skills/ instead of ~/.claude/skills/.

Once linked, the agent will automatically invoke zhub when you ask it to manage agents, skills, or providers.

Security

• Tokens stored in ~/.zhub/config.yaml with file permission 0600
• Provider apiKey never shown in plaintext — masked to <hidden> in all output formats
• MCP headers never shown in plaintext — masked to <hidden> in mcp get output (empty header values remain empty)
• All commands non-interactive — designed for automation, no prompts or TUI

Multiple Profiles

zhub login --url https://staging.hub... --token cli_yyy --profile staging
zhub --profile staging agent list

Development

cd cli
bun install
bun test                    # 90 tests
bunx tsc --noEmit           # type check
bun run src/index.ts <cmd>  # run without install

Publish: bun publish (requires npm login + 2FA)

License

AGPL-3.0