Building an AI agent that can actually do things — send Slack messages, create Jira tickets, update spreadsheets — means solving authentication for every single app. OAuth flows, token refresh, API quirks, rate limits. For one app, it’s annoying. For ten apps, it’s a full-time job.

The Zapier SDK (open beta, April 2026) solves this by giving your agent authenticated access to 9,000+ app integrations through one interface. Zapier handles auth, token refresh, retries, and API differences. Your agent handles the logic.

SDK vs MCP vs Zapier Agents

Zapier now has three ways to connect AI to apps. They serve different purposes:

	Zapier SDK	Zapier MCP	Zapier Agents
For	Coding agents and developers	Chat agents (Claude, GPT)	No-code users
Access	Any API call, in code	Curated menu of pre-built actions	Visual builder
Auth	Programmatic (client credentials)	Browser-based	Browser-based
Use when	You need loops, conditionals, error handling	You want quick tool access in chat	You want no-code automation

Most teams end up using SDK + MCP together. MCP for conversational interfaces, SDK for production code.

Setup

# Requires Node.js 20+
npm init -y
npm install @zapier/zapier-sdk
npm install -D @zapier/zapier-sdk-cli @types/node typescript

# Login (opens browser)
npx zapier-sdk login

# Check your connected apps
npx zapier-sdk list-connections --owner me

The CLI is useful for exploration. Use it to discover what actions an app supports:

# See all available actions for Slack
npx zapier-sdk list-actions slack

# See what inputs an action needs
npx zapier-sdk describe-action slack write channel_message

# Run an action directly
npx zapier-sdk run-action slack write channel_message \
  --connection-id YOUR_ID \
  --inputs '{"channel":"#general","text":"Hello from the SDK"}'

TypeScript SDK usage

The SDK is type-safe. Every app and action has generated types:

import { createZapierSdk } from "@zapier/zapier-sdk";

const zapier = createZapierSdk();

// Find the user's Slack connection
const { data: slackConn } = await zapier.findFirstConnection({
  appKey: "slack",
  owner: "me",
  isExpired: false,
});

// Bind it once, use everywhere
const slack = zapier.apps.slack({ connectionId: slackConn.id });

// Send a message
await slack.write.channel_message({
  inputs: {
    channel: "#deployments",
    text: "Deploy complete. All tests passing.",
  },
});

The pattern is consistent across all 9,000+ apps: zapier.apps.APP_KEY({ connectionId }).ACTION_TYPE.ACTION_NAME({ inputs }).

Action types:

• search — find data (search users, find events)
• write — create or update (send message, create ticket)
• read — list data (list channels, get records)

Real example: AI agent with tool access

Here’s how you’d give an AI agent access to Slack and GitHub through the Zapier SDK:

import { createZapierSdk } from "@zapier/zapier-sdk";

const zapier = createZapierSdk();

// These become tools your agent can call
async function sendSlackMessage(channel: string, text: string) {
  const { data: conn } = await zapier.findFirstConnection({
    appKey: "slack", owner: "me", isExpired: false,
  });
  const slack = zapier.apps.slack({ connectionId: conn.id });
  return slack.write.channel_message({ inputs: { channel, text } });
}

async function createGithubIssue(repo: string, title: string, body: string) {
  const { data: conn } = await zapier.findFirstConnection({
    appKey: "github", owner: "me", isExpired: false,
  });
  const github = zapier.apps.github({ connectionId: conn.id });
  return github.write.create_issue({
    inputs: { repo, title, body },
  });
}

// Register these as tools in your agent framework
// Works with OpenAI Agents SDK, LangChain, CrewAI, etc.

This is the key advantage: your agent gets authenticated access to thousands of apps without you building a single OAuth flow.

Direct API calls with fetch

For actions that go beyond pre-built connectors, the SDK provides authenticated fetch:

const { data: conn } = await zapier.findFirstConnection({
  appKey: "github", owner: "me", isExpired: false,
});

// Raw API call — Zapier handles auth headers
const response = await zapier.fetch(conn.id, {
  url: "https://api.github.com/repos/owner/repo/pulls",
  method: "GET",
});

This gives you full API access with Zapier managing the credentials. Note: direct API calls are not yet covered by Zapier’s governance policies (pre-built actions are).

Governance and security

If your organization has restricted specific apps or actions in Zapier, those policies apply automatically to SDK traffic using pre-built actions. This matters for enterprise teams where AI agent security is a concern.

Current limitations (open beta):

• Enterprise and Team plans are off by default (contact Zapier for opt-in)
• Direct API calls via .fetch() bypass governance policies (fix coming)
• Triggers API (real-time events) targeting May 2026
• Agent approval flow (user reviews before agent acts) coming soon

SDK vs building it yourself

	Zapier SDK	DIY
Time to first integration	Minutes	Hours to days
OAuth implementation	Handled	You build it
Token refresh	Automatic	You handle it
Number of apps	9,000+	One at a time
API quirks	Abstracted	You debug them
Cost	Free (beta), paid later	Free (your time)
Vendor lock-in	Yes (Zapier)	No

The trade-off is clear: speed and breadth vs independence. For most AI agent projects, the SDK is worth it for the auth layer alone.

Combining with agent frameworks

The Zapier SDK works with any agent framework:

• OpenAI Agents SDK: Register Zapier actions as @function_tool tools
• LangChain: Wrap as LangChain tools with @tool decorator
• Claude Code: Use via MCP for conversational access
• CrewAI: Register as crew tools for multi-agent workflows

The SDK handles the “connect to apps” problem. Your framework handles the “think and decide” problem. Together, they make agents that can actually do useful work.

Related: OpenAI Agents SDK Guide · How to Build an AI Agent · Best AI Agent Frameworks · What is MCP? · Agent Orchestration Patterns · AI Agent Security · Agent vs Workflow