Deutsch | Español | français | 日本語 | 한국어 | Português | Русский | 中文
📖 Full docs: docs.zibby.app · Get Started · Concepts · Designing agents · CLI Reference · Cloud
The skill layer for @zibby/agent-workflow. Built-in skill definitions that give a workflow node the tools it needs — function tools, MCP servers, browser, issue trackers, memory. Vendor-neutral, JavaScript-first.
@zibby/skills is the batteries-included companion to @zibby/agent-workflow (npm) — "Graph-based AI agent workflow orchestration." The workflow engine ships zero skills on purpose; this package is where the built-in ones live.
A skill is the contract between a workflow node and a tool. It tells the engine what the tool does, how to start it, and what it needs. The engine never hardcodes any skill by name — it reads the skill definition and wires things up generically for both Claude and Cursor agents.
@zibby/agent-workflow node @zibby/skills
────────────────────────── ─────────────
skills: ['add'] ──► getSkill('add')
│
▼
skill.resolve() → { command, args, env }
│
┌────────┴────────┐
▼ ▼
Claude SDK Cursor CLI
You don't use @zibby/skills on its own — it plugs into @zibby/agent-workflow. A node names the skills it wants in its skills: array, and the workflow engine resolves them at run time:
npm install @zibby/agent-workflow @zibby/skills// 1. Import the package to register all built-in skills
import '@zibby/skills';Define a skill once…
import { skill } from '@zibby/skills';
export const add = skill('add', {
description: 'Add two numbers',
input: { a: 'number', b: 'number' },
handler: async ({ a, b }) => ({ result: a + b })
});…and an @zibby/agent-workflow node requests it by id:
// Used by an @zibby/agent-workflow node
export const mathNode = {
name: 'do_math',
skills: ['add'],
prompt: (state) => `Add ${state.a} and ${state.b}`,
};When the engine runs do_math, it sees skills: ['add'], looks the skill up, calls resolve(), and hands the resulting tool to whichever agent runs the node. See @zibby/agent-workflow for how nodes, graphs, and state fit together.
npm install @zibby/skillsImport the package to register all built-in skills:
import '@zibby/skills';One function to create any skill. Auto-detects the type and auto-registers.
One skill = one tool. Flat, no nesting.
import { skill } from '@zibby/skills';
export const add = skill('add', {
description: 'Add two numbers',
input: { a: 'number', b: 'number' },
handler: async ({ a, b }) => ({ result: a + b })
});Use it in an @zibby/agent-workflow node:
export const mathNode = {
name: 'do_math',
skills: ['add'],
prompt: (state) => `Add ${state.a} and ${state.b}`,
};For wrapping existing MCP server packages:
import { skill } from '@zibby/skills';
export const linear = skill('linear', {
envKeys: ['LINEAR_API_KEY'],
description: 'Linear issue tracker',
resolve() {
if (!process.env.LINEAR_API_KEY) return null;
return {
command: 'npx',
args: ['-y', '@modelcontextprotocol/server-linear'],
env: { LINEAR_API_KEY: process.env.LINEAR_API_KEY },
};
}
});| ID | Server | MCP Package |
|---|---|---|
browser |
playwright |
@zibby/mcp-browser / @playwright/mcp |
jira |
jira |
@zibby/mcp-jira |
github |
github |
@modelcontextprotocol/server-github |
slack |
slack |
@modelcontextprotocol/server-slack |
skill(id, { description, input, handler })id— Unique skill identifier (used inskills: ['add'])description— What the tool does (shown to the LLM)input— Parameter definitions:
{
param: { type: 'string' }, // full form
other: 'number', // shorthand
optional: { type: 'string', required: false },
}handler— The function that runs when the tool is called:
handler: async ({ param, other }) => {
return { result: 'something' }; // any JSON-serializable value
}- Must be
async(or return a Promise) - Receives one object argument with the input parameters
- Must return a JSON-serializable value
- Has full access to imports, closures, and the module scope
- Runs in a child process (the function bridge)
import { skill } from '@zibby/skills';
export const fetchUrl = skill('fetch_url', {
description: 'Fetch a URL and return the response body',
input: { url: 'string' },
handler: async ({ url }) => {
const res = await fetch(url);
return { status: res.status, body: await res.text() };
}
});
export const healthCheck = skill('health_check', {
description: 'Check if the service is running',
handler: async () => ({ status: 'ok', timestamp: Date.now() })
});skill(id, config)Config object:
| Property | Required | Description |
|---|---|---|
resolve(options) |
Yes | Returns { command, args, env } or null |
serverName |
No | MCP server name (defaults to id) |
allowedTools |
No | Tool patterns (defaults to ['mcp__<serverName>__*']) |
envKeys |
No | Env vars the skill needs |
description |
No | Human-readable description |
tools |
No | Tool schemas for compile-time validation |
cursorKey |
No | Override key in ~/.cursor/mcp.json |
sessionEnvKey |
No | Env var for session artifact paths (Cursor only) |
import { skill } from '@zibby/skills';
import { createRequire } from 'module';
const _require = createRequire(import.meta.url);
export const database = skill('database', {
envKeys: ['DATABASE_URL'],
description: 'Database query MCP server',
resolve({ sessionPath } = {}) {
let bin;
try { bin = _require.resolve('@myorg/mcp-database/server.js'); }
catch { bin = null; }
if (bin) {
return {
command: 'node',
args: [bin, '--read-only'],
env: { DATABASE_URL: process.env.DATABASE_URL },
};
}
return {
command: 'npx',
args: ['-y', '@myorg/mcp-database', '--read-only'],
env: { DATABASE_URL: process.env.DATABASE_URL },
};
}
}); Node definition Skill definition Agent strategy
───────────── ──────────────── ──────────────
skills: ['add'] ──► getSkill('add') ──► strategy-specific setup
│
▼
skill.resolve()
│
▼
{ command, args, env }
│
┌────────┴────────┐
▼ ▼
Claude SDK Cursor CLI
────────── ──────────
In-memory Writes to
mcpServers ~/.cursor/mcp.json
param to before spawning
query() `agent` CLI
Claude: The SDK receives mcpServers as a parameter. It spawns the MCP server as a child process, connects via stdio, routes tool calls through it.
Cursor: The engine writes ~/.cursor/mcp.json to disk before spawning the agent CLI. Cursor reads that file and manages MCP servers itself.
The strategies never reference any skill by name. They loop over the skill definitions and call resolve() on each.
import {
skill, // Unified factory — auto-detects type, auto-registers
registerSkill, // Register a raw skill definition
getSkill, // Get a skill by ID
hasSkill, // Check if a skill is registered
getAllSkills, // Get all registered skills (Map)
listSkillIds, // Get array of registered skill IDs
SKILLS, // Built-in skill ID constants
} from '@zibby/skills';| Package | What it adds |
|---|---|
@zibby/agent-workflow |
The graph engine. Skills here plug into its nodes. |
@zibby/cli |
zibby command — scaffold, dev server, deploy, trigger, logs. |
@zibby/core |
Built-in agent strategies (Claude / Cursor / Codex / Gemini / OpenAI Assistant), MCP client, runtime. |
MIT