Zapier MCP Server: Connect Agents to Apps and the Web

TL;DR

• Zapier MCP connects your AI agent to thousands of apps. Through one connection, it can update a CRM, send a Slack message, or add a row to a spreadsheet.
• Works in both directions. An AI agent can use Zapier to act on apps, while a Zap can call an external MCP server when it needs something Zapier doesn't provide.
• The Zapier MCP server doesn't provide live web access. Pair it with Decodo when your workflow needs to scrape webpages or retrieve search results.
• You can control what the agent can use by choosing its actions and letting Zapier discover the tools it needs as it works.
• The trade-off is breadth over depth, since each successful tool call uses two Zapier tasks and native MCP servers offer deeper, service-specific tools.

What is the Zapier MCP server, and how does it connect AI agents to apps?

The Zapier MCP server is Zapier's implementation of the Model Context Protocol. It connects an AI client to Zapier's app catalog through a single remote MCP connection, so you don't need a separate integration for every app. Zapier handles the app connections, credentials, and rate limits, while the AI client discovers and calls the actions you allow it to use.

Core features of Zapier MCP

• Single connect URL. Connect your AI client to Zapier MCP using one remote URL instead of setting up a separate connection for every app.
• OAuth. Sign in to your Zapier account once to let Zapier use the app accounts you've already connected.
• Dynamic tool discovery. Let the AI find the Zapier action it needs when it needs it instead of loading thousands of actions into its context at once.
• Per-action configuration. Decide which actions the AI can use and control which values it can provide when those actions run.
• Per-app MCP pages. Connect to a specific app or set of actions when you want to give the AI a narrower set of tools.
• Zapier Agents. Use Zapier's no-code agent builder to create agents that can use these actions without building the agent yourself.

How Zapier MCP connects to your AI client

The features above work through a hosted remote MCP server. Unlike local MCP servers, which run as a Node.js process on your machine and talk to the client over stdio, Zapier MCP needs no local install, runtime, or dependency updates. You get the MCP connection URL from Zapier, add it as a remote MCP server in your AI client, and complete the OAuth flow to connect your Zapier account.

Once connected, the client doesn't load every available action at once. Instead, Zapier uses meta-tools to help it find and run the actions it needs when they're needed, to keep unnecessary tool definitions out of the client's context.

How actions and plugins work

With your AI client connected to the Zapier MCP server, you can decide what it can actually do. You can connect directly through the hosted MCP server, or use a Zapier MCP plugin if your client offers one.

A Zapier MCP plugin is a client-specific package that sits on top of that hosted connection. It adds things such as guided onboarding, a first-use demo, and skills that help the AI use Zapier. The underlying MCP server remains the same. If you don't use a plugin, you can connect directly using Zapier's MCP server URL. For a broader look at setting up MCP servers, see how to set up an MCP server.

Zapier MCP servers also run in one of two modes. Agentic mode is the default for new servers. Zapier pre-enables actions for the apps you've already connected, and the agent searches for, enables, and runs anything else it needs as requests come up, using a fixed set of meta-tools such as discover_zapier_actions and enable_zapier_action. Managed mode works the other way around. You pick the actions in advance at mcp.zapier.com, each one becomes a dedicated tool, and the agent can't add new ones. Managed mode also lets you lock specific field values per tool, which suits repeatable or automated tasks. Saved skills are only available in agentic mode.

Whichever connection you use, the next step is to decide how much access to give the AI. Start with a few basic guardrails:

• Keep a human in the loop for actions with meaningful side effects. Sending a message, creating a record, or modifying data should be reviewed before they run.
• Give the AI only the actions it needs. Don't expose every action in a connected app just because it's available.
• Start with a low-stakes test. Run a read action first, or a write that's easy to undo, such as adding a row to a test spreadsheet.

Then configure each action the AI can use:

• Enable the action. Choose the app and operation you want the AI to perform. In agentic mode, the agent can do this for you as it works. In managed mode, you set the list yourself.
• Authenticate the app. Connect the app account the action will use. Each app needs its own account connection, even when the actions are used through the same Zapier MCP connection.
• Configure the fields. In managed mode, decide which values should stay fixed and which the AI should provide at call time. Use static values for things that shouldn't change, such as a Slack channel or Google Sheet. Use AI-filled fields for changing content, such as the message you want to send. If a required value is missing or invalid for the connected app, the action won't run.

Once the action is configured, the AI client can discover it through Zapier's MCP tools and call it when a user's request matches. Zapier runs the action against the connected app and returns the result to the AI.

A successful Zapier MCP tool call uses two Zapier tasks. Failed tool calls don't consume tasks, but repeated successful tests can still use your task allowance quickly while you're debugging.

Two directions: Agents into Zapier, Zaps into MCP servers

So far, we've looked at how an AI agent connects to Zapier MCP and uses its actions. But Zapier can also work in the opposite direction: a Zap can call an external MCP server when the workflow needs a capability that isn't available through a Zapier app. This is different from A2A, where it handles communication between AI agents, while MCP connects an AI agent to external tools and systems.

Direction A: Agent → apps

The AI agent starts the interaction. You ask it to do something, and it uses Zapier MCP to take that action in an app such as Slack, Google Sheets, or Salesforce. In this setup, the agent decides what action to take, while Zapier carries it out in the connected app.

Direction B: Zap → MCP tools

The Zap starts the workflow. When one of its steps needs information or a capability that isn't available through a Zapier app, the Zap can call an external MCP server instead. For example, a Zap can call an MCP server to retrieve the latest details from a product page, then pass that information to the next step in the workflow.

Which one should you use? Choose Zapier MCP when the interaction starts with an AI agent. Choose MCP Client by Zapier when an event triggers a Zap that needs to call an external MCP server.

Getting started: Connecting Zapier MCP to your AI client

The quickest way in is to add Zapier as a connector inside your AI client and sign in when prompted. Zapier creates and configures your server during that sign-in, so you don't need to set one up at mcp.zapier.com first. Clients that ask for a URL use Zapier's connect URL, https://mcp.zapier.com/api/v1/connect, added as a remote HTTP server.

Once you're signed in, Zapier pre-enables actions for the apps already connected to your Zapier account. To get set up from there, ask your agent to load the zapier:onboarding skill. It checks what's enabled, adds apps if nothing is, gives you the sign-in links in one batch, and then runs a quick demo across a few of your apps.

• For Zapier MCP in Claude, add Zapier from the connectors directory and sign in. Claude Code uses the same connection; follow the Claude MCP setup guide for client-specific steps.
• For Cursor, add the Zapier MCP connection URL as a remote MCP server and complete authentication. See how to add an MCP server to Cursor.
• For ChatGPT, full MCP support, including write and modify actions, is in beta for Business, Enterprise, and Edu workspaces, on the web only. Pro users can connect MCP apps with read/fetch permissions through developer mode. See OpenAI's developer mode article for plan details and where to find the setting.
• Replit and other MCP-compatible clients can use the same Zapier MCP URL. Only the location of the MCP settings and the OAuth flow vary by client.

Once connected, test the setup with a simple action before giving the agent access to anything consequential. For example, have it create a test row in Google Sheets and confirm that it works before moving on to actions that send messages, modify important records, or trigger other side effects.

Use cases: What teams run through Zapier MCP

When you want an AI agent to work with the apps your team already uses, these are the kinds of tasks you can hand off.

Direction A: Agent → apps

If you're building agents with AI agent development tools, Zapier MCP gives them a way to take action in the apps your team already uses. 

• Manage your inbox or calendar. "Find a free slot tomorrow afternoon and schedule a meeting with Sarah."
• Update a CRM after a call. "Add these call notes to the Acme opportunity and create a follow-up task for Friday."
• Create a ticket from your IDE. "Create a Jira ticket for this bug and assign it to the payments team."
• Log analysis results in a spreadsheet. "Add these findings and metrics to the Q3 analysis sheet."

What fits less well are long, unsupervised chains of actions, where every extra step is another tool call, another two tasks, and another chance for the agent to pick the wrong action with nobody checking. Fixed, repeatable flows like these usually run better as a Zap, which is where Direction B comes in.

Direction B: Zap → MCP tools

A Zap starts the workflow, then calls an external MCP server when it needs data or capabilities that aren't available through a Zapier app.

• Generate scheduled research digests using current web data.
• Enrich new leads with company information when a new row is added.
• Monitor competitor websites and pricing for changes.
• Check content before publishing using live pages, search results, or other web data.

These web-facing workflows need a source for the data itself, which is where an external MCP server such as Decodo comes in. Decodo gives the workflow access to live web data to perform tasks such as researching competitors, tracking pricing changes, and checking published content.

Adding live web data: Decodo with Zapier MCP

Zapier MCP gives AI agents access to thousands of apps, but it doesn't scrape the open web on its own. If an agent needs information from a webpage or search results, it needs a way to retrieve that data.

That's where Decodo comes in. Decodo provides the web data layer, retrieving live web data and making it available to the agent or workflow.

Consider a simple example. You want to monitor a competitor's pricing page, detect when the price changes, and get the updates on Slack.

There are three ways to set this up:

1. Use Decodo as a Zapier action through Zapier MCP.
2. Connect Decodo MCP directly alongside Zapier MCP.
3. Call Decodo MCP from a Zap using MCP Client by Zapier.

Which setup you choose depends on how your workflow is built today, where you want the web data connection to happen, and how much control you need over Decodo's tools.

Pattern 1: Use Decodo as a Zapier action through Zapier MCP

In this setup, the agent reaches Decodo through Zapier MCP. Decodo's actions show up alongside those of every other connected app, and the agent can discover the one it needs and call it without adding another MCP server.

For the pricing example we mentioned, the agent uses the Decodo action to retrieve the competitor's pricing page. It then checks the result for a pricing change and uses another Zapier action to send the update to Slack.

Competitor pricing page → Decodo action through Zapier MCP → Agent → Slack

Decodo's Zapier integration provides three actions, all backed by the Web Data API.

• Fetch URL. Scrapes one page and returns Markdown, HTML, or a PNG screenshot.
• Fetch URL List. Scrapes up to 25 URLs per run and returns each page's content, plus counts and failure details.
• Run Search. Runs a SERP or e-commerce search and returns parsed results or page content.

Decodo also has its own page on Zapier MCP, which lists the same three actions. To connect it, use a Web Data API key from your subscription on the Decodo dashboard.

Here's the flow we ran with Claude to scrape the first three pages of a demo shop and drop every product into a Google Sheet. The same steps apply to the Slack pricing flow; only the destination app changes.

1. Connect Zapier MCP to your agent. In Claude, that's the Zapier connector. Zapier pre-enables actions for apps you've already connected. Our test account had none, so we added them as we went.
2. Add Decodo. Ask the agent to add Decodo, and it finds the app with discover_zapier_actions and enables it with enable_zapier_action. This adds all three actions (scrape_url, scrape_url_list, and run_search) in one go.
3. Authenticate. Zapier returns a sign-in link for Decodo. Open it and paste your Web Data API key.
4. Set the default connection. If an action reports that no default connection is set, ask the agent to fix it. It lists the app's connections with list_zapier_connections and sets yours as the default with manage_zapier_connections.
5. Add the destination app. For this run, that was Google Sheets, enabled and authenticated the same way.
6. Scrape. The agent inspected the action schema with inspect_zapier_actions, then called Fetch URL List with the three page URLs and Markdown output turned on.
7. Shape and deliver. The agent parsed the Markdown into product name, price, product URL, and image URL. It then created a new spreadsheet with headers and wrote all 48 products in a single Create Multiple Spreadsheet Rows call.

Beyond two quick sign-ins, the whole run took three short prompts. Everything in between ran through Zapier MCP.

This setup fits if your agent already uses Zapier MCP and Decodo's three actions cover what the workflow needs. The scraping and the delivery (Sheets, Slack, email, a CRM) live behind one connection, and you manage the Decodo key in Zapier instead of running Decodo MCP separately.

Pattern 2: Connect Decodo MCP alongside Zapier MCP

You don't need the Decodo Zapier integration to use the two together. An agent can connect directly to the Decodo MCP server alongside Zapier MCP.

The two servers handle different parts of the workflow. Decodo MCP gives the agent tools for retrieving live web data through Decodo's Web Data API, while Zapier MCP gives it actions for working with apps such as Slack and Google Sheets, including the one where previous results are stored.

For the pricing workflow, the process looks like this:

1. Retrieve the pricing page with Decodo MCP. The agent calls scrape_as_markdown with the competitor's pricing page URL and gets back clean Markdown, ready to parse.
2. Load the previous prices through Zapier MCP. The agent reads the last stored results from wherever you keep them, such as a Google Sheet.
3. Check whether the price changed. The agent extracts the current prices from the Markdown and compares them with the stored ones.
4. Send the update through Zapier MCP. If it finds a change, the agent calls the relevant Slack action to post the new pricing, then writes the latest results back to storage for the next run.

The complete flow looks like this:

Competitor pricing page → Decodo MCP → Agent ⇄ stored prices (via Zapier MCP) → Zapier MCP → Slack

Because the agent connects to Decodo MCP directly, it can use the full set of tools the server exposes, not only the three Decodo actions available through Zapier. Beyond scrape_as_markdown, that includes screenshots, site mapping, multi-page crawling, and tools for querying Google AI Mode, ChatGPT, and Perplexity. For long pricing pages, scrape_as_markdown also takes a tokenLimit parameter, so you can control how much content comes back.

There's another practical benefit: separation. Each server stays focused on its own job, with Decodo handling web data and Zapier handling app actions. You decide which Decodo tools and which Zapier actions each agent gets, which keeps tool selection simple as workflows grow.

To connect Decodo MCP, use the hosted server at https://mcp.decodo.com/mcp with your basic auth token in an Authorization header, or run it locally with npx -y @decodo/mcp-server and the token set as SCRAPER_API_TOKEN. The setup documentation covers each supported AI client.

Pattern 3: Call Decodo MCP from a Zap

The first two patterns start with an AI agent. Not every workflow starts there, though. Sometimes it already lives in Zapier.

Say you want to check a competitor's pricing page every morning instead of waiting for an agent to decide when to run. In that case, a Zap triggers the workflow and calls Decodo MCP when it needs web data.

MCP Client by Zapier lets a Zap connect to any remote MCP server and call its tools as a regular step. It’s currently in beta, so expect some changes as Zapier develops it. It offers two actions and two triggers.

• Run Read-Only Tool. Calls a read-only tool and returns one or more results.
• Run tool. Calls any tool and returns one result.
• New Tool Result. A trigger that fires when a read-only tool returns a new result.
• New Tool. A trigger that fires when the server adds a new tool.

For the pricing example, a scheduled Zap could look like this:

1. Schedule by Zapier starts the Zap every morning.
2. MCP Client by Zapier runs scrape_as_markdown on the competitor's pricing page using Run Read-Only Tool.
3. A storage step, such as Google Sheets or Storage by Zapier, holds the previous price for comparison.
4. Filter by Zapier lets the Zap continue only when the price has changed.
5. Slack posts the update.

Scheduled Zap → MCP Client by Zapier → Decodo MCP → Filter → Slack

The setup takes a couple of minutes with the MCP Client by Zapier app:

• Server URL. Enter https://mcp.decodo.com/mcp.
• Transport. Select Streamable HTTP.
• OAuth. Set Use OAuth to No, since Decodo MCP authenticates with a token.
• Authorization. Add your Web Data API basic auth token from the Decodo dashboard.

Once connected, Decodo's tools appear in the Tool dropdown, including scrape_as_markdown, screenshot, google_search, google_ai_mode, and the Amazon, Walmart, Target, TikTok Shop, and YouTube tools. Most are marked read-only, so they're available in Run Read-Only Tool as well as Run tool.

In our test, Run Read-Only Tool called scrape_as_markdown on a demo shop page and returned the full product listing as Markdown on the first attempt. The response comes back as a list of content items, with the page Markdown in the second one. To pass only the page content to later steps, set Filter result to a JSONata query such as {"page": content[1].text}. The tokenLimit field caps how much content comes back, which keeps the payload small for steps that only need part of the page.

To keep the tool list focused, you can add a toolsets parameter to the server URL. For example, https://mcp.decodo.com/mcp?toolsets=web exposes only the web scraping tools.

Use this pattern when the workflow already starts with a Zap trigger, and Decodo only needs to provide web data at one or more steps.

The three patterns differ mainly in where the workflow starts and where Decodo connects to it.

Three ways to give a Zapier workflow live web data

For more on using Decodo MCP with AI agent workflows, see the n8n and Decodo MCP orchestration guide for a similar setup outside Zapier.

Zapier MCP vs. native MCP servers: When each wins

The three patterns above show different ways to connect an agent or Zap to live web data. But the same question comes up when connecting an AI agent to other services: if the agent can connect directly to an app through its native MCP server, why use Zapier MCP at all?

It comes down to what you need from the connection. Zapier MCP gives an agent broad access to apps that may not have their own MCP server. Native MCP servers give it more direct access to the tools and schemas built for a specific service.

That creates a trade-off between reach and depth. Zapier MCP connects an AI client to 9,000+ apps and 40,000+ actions, while Zapier handles the app connections, credentials, and rate limits. A native MCP server can instead expose specialized tools and parameters designed specifically for one service.

Reach vs. depth

Think about an agent that needs to work across Slack, Salesforce, Google Calendar, and an internal tool.

If those services don't all offer native MCP servers, connecting each one directly isn't an option. Zapier MCP can give the agent access to its Zapier actions through one MCP connection.

Native MCP servers become more useful when the agent needs deeper access to a particular service. A native server can expose tools and parameters specifically designed for that service instead of relying on the actions available through a general automation platform.

So the trade-off is:

• Zapier MCP. Broader app reach.
• Native MCP. Deeper, service-specific access.

What does this mean for context cost?

Connecting multiple native MCP servers can create another problem: tool definitions take up context.

Anthropic gives a useful example. Five MCP servers with 58 tools consumed about 55,000 tokens just for their tool definitions. In another example, loading 50+ MCP tools upfront used about 72,000 tokens before the agent started working.

But that doesn't mean Zapier MCP automatically uses less context than native MCP servers. How the AI client discovers and loads tools also matters.

Modern clients are increasingly moving toward on-demand tool discovery. Cursor, for example, now loads MCP tool details only when they're needed and reported a 46.9% reduction in total agent tokens in an A/B test of runs that called MCP tools. Anthropic's Tool Search takes a similar approach by loading relevant tools when they're needed.

So context cost depends on both the MCP setup and how the AI client handles tool discovery. A native MCP setup doesn't automatically mean higher context usage, just as Zapier MCP doesn't automatically mean lower usage.

Where can things fail?

Zapier MCP adds one more layer between the agent and the app, so there's one more place where a request can fail. A native MCP setup gives you a relatively direct path:

Agent → Native MCP server → Service

With Zapier MCP, the path has another layer:

Agent → Zapier MCP → Zapier integration → Service

That extra layer can be useful because Zapier manages connections, credentials, and rate limits. But it also means there's another place where a request can fail, or configuration can drift. A failed action may require you to check the Zapier connection or action configuration before you can determine whether the underlying app caused the problem. Zapier also lets you scope which tools are available to the AI, which adds another configuration layer to maintain.

Native servers can make that path easier to inspect when you are working deeply with one service. But you then take responsibility for connecting and managing each server separately. If you're weighing which native servers are worth that effort, our roundup of the top MCP servers for AI workflows covers options for web research, database access, and development tools.

When should you use Zapier MCP?

Use Zapier MCP when your agent needs to reach many apps through one connection, and use a native MCP server when it needs deep, specialized access to one service. There isn't a single alternative that covers the same use case in exactly the same way.

If you need an agent to reach a large number of apps, Zapier MCP solves a problem that individual native MCP servers cannot solve on their own. Its value is the breadth of its app catalog.

If you need deep, specialized access to a particular service, a native MCP server may make more sense.

If what you actually need is event-driven automation, the comparison changes. n8n and Make are built around workflows that start with an event or schedule, run a sequence of steps, and pass data between services. That's a different model from giving an AI agent a large catalog of tools it can discover and call. 

Pricing, limits, and where to get help

Zapier MCP pricing follows your existing Zapier plan rather than requiring a separate MCP subscription. MCP is available to all accounts, and each successful MCP tool call uses two tasks from your account's shared task allowance. So, 10 successful MCP calls use 20 tasks, and those tasks come from the same pool used by your other Zapier usage.

Usage limits

Your available MCP usage depends on the task allowance in your Zapier plan. Current task tiers range from 100 to 2 million tasks, and MCP shares that allowance with other Zapier usage. If you reach your task limit, MCP calls stop until the allowance resets or you upgrade.

Setting up actions

You don't need to configure every action upfront. Zapier MCP can discover and enable tools during a conversation. You still need to connect the relevant app accounts and control which tools the agent can use when tighter access is required.

Reliability and security

Zapier handles app connections, credentials, and rate limits for MCP. On the usage side, the task allowance is the only limit, with no cap on calls per session.

Keep the agent's tool access as narrow as the workflow allows. For actions with meaningful side effects, use human approval where appropriate rather than giving the agent unrestricted access.

Where to get help

• MCP documentation. Get started, connect supported clients, and learn how tools work.
• Per-app MCP pages. Check the tools available for individual apps.
• Zapier MCP GitHub repository. Find the official MCP resources and client setup files.
• Zapier Support and Zapier Community. Get help with setup and troubleshooting.
• Zapier Status. Check for current service issues.

For the Decodo side, see the Decodo MCP documentation for setup and troubleshooting.

Final thoughts

Zapier MCP connects your AI agent to thousands of apps, so it can use the tools and services you already work with without building separate integrations for each one. That gives your agent a way to take action across your SaaS stack.

But apps aren't the whole picture. If you want your AI agent to check a competitor's pricing page, monitor search results, or pull information from a website, it needs a way to retrieve live data from the web. That's where a web scraping tool like Decodo comes in.

Decodo lets you add live web data to your Zapier-based AI workflows in three ways. For conversation-triggered workflows, you can connect the agent to both Decodo MCP and Zapier MCP. For event-triggered workflows, you can use a Zap with MCP Client by Zapier to call Decodo MCP when the workflow needs fresh web data. If your agent already uses Zapier MCP, you can add Decodo as a Zapier app and keep everything in one connection.