Zapier MCP: setup, pricing, tools and limits

Short answer

Zapier MCP is Zapier's hosted MCP server: one address, https://mcp.zapier.com/api/v1/connect, that lets an AI agent run actions in 9,000+ apps through your Zapier account, like finding a contact in HubSpot, adding a row to a sheet or posting to Slack. Add it to Claude, ChatGPT, Codex, Cursor or any MCP client and sign in to Zapier. Each successful tool call uses 2 tasks from your Zapier plan, so the Free plan's 100 tasks cover about 50 calls a month. In Claude Code:

claude mcp add --transport http zapier https://mcp.zapier.com/api/v1/connect

Then run /mcp to sign in. Zapier covers apps that have no MCP server of their own. For apps that do, like Gmail, Slack or Notion, their own servers give agents more tools. Upsy holds both, with a level per agent.

Zapier's MCP server

Zapier's MCP docs and help article have the details. Zapier MCP replaced Zapier's AI Actions and its older per-server URLs.

Servers and modes

A Zapier MCP server holds the actions your agent may use, their settings, and a History of every call. Signing in from a client creates your default server, and since October 2026 several clients can share one server. Each server runs in one of two modes, set under Settings at mcp.zapier.com:

• Agentic (the default): your client gets 16 fixed tools and finds, turns on and runs actions as it goes. Zapier pre-enables actions for apps you've already connected in Zapier.
• Managed: you pick each action at mcp.zapier.com, and each one becomes its own tool. The agent can't add more.

In both modes, you can fix a field's value for an action, so the agent can't change it, such as always posting to one Slack channel.

What it costs

• Tasks are shared with your Zaps. Searching and updating 10 records is 11 calls, or 22 tasks (Zapier's usage page).
• At the limit, paid plans with pay-per-task keep running at a higher task price, up to three times the plan's allowance. Otherwise calls fail until the month resets.
• From December 1, 2026, calls to Zapier's own free apps, such as Tables and Forms, cost no tasks, and AI by Zapier calls are billed at its own price.
• There's no separate rate limit published for MCP; your plan's tasks are the cap.

Connect it to your client

Claude. Add Zapier from Claude's connector directory (Customize › Connectors) and sign in. It works on every Claude plan; on Team and Enterprise, an Owner turns it on first.

Claude Code. Run the command above, or install Zapier's plugin with /plugin install zapier@claude-plugins-official.

ChatGPT. Open Plugins, click Add, then Create custom MCP server. Name it Zapier, paste the address, keep OAuth and click Create as a plugin. On Business and Enterprise, use the Zapier app your workspace installed. See how to add a custom MCP server to ChatGPT.

Codex. Run codex mcp add zapier --url https://mcp.zapier.com/api/v1/connect, then codex mcp login zapier if no browser opens.

Cursor. Install Zapier from Cursor's marketplace. Update Cursor first; older versions fail at sign-in.

VS Code. Run MCP: Open User Configuration and add {"servers":{"Zapier":{"url":"https://mcp.zapier.com/api/v1/connect"}}}.

Gemini CLI. Run gemini extensions install https://github.com/zapier/zapier-mcp, then /mcp auth zapier.

Your own code. Generate a connection token under the server's Settings › API token and send it as Authorization: Bearer. It's shown once and tied to one server. Zapier's client list covers about 40 more, including Windsurf, Grok and Slackbot.

Its tools

In agentic mode, every server has the same 16 tools (Zapier's list):

• Every app action runs through two tools. Reads and searches go through execute_zapier_read_action, everything else through execute_zapier_write_action, including deletes where an app offers them.
• "Find or Create" actions in apps like Salesforce and HubSpot search first and create a record when nothing matches. Zapier doesn't say which of the two tools runs them, so don't count on the read tool alone to keep an agent from creating records.
• Skills are saved Markdown instructions, like "how we triage support email", that the agent loads when it needs them.
• Some accounts also get write_code_action, which writes and runs code against an app's API when no action fits. With early access, about 25 more tools build Next Gen Zaps.

Where Zapier MCP stops

• Actions, not every API. Zapier offers a fixed set of actions for each app, often fewer than the app's own MCP server, which may also read whole threads, files or boards.
• No events. Tools run when the agent calls them. Reacting to new emails or form entries takes a Zap, or a Next Gen Zap built in early access.
• Every call costs tasks, including searches the agent runs just to look around.
• One level per server. Clients on the same server share its actions. Keeping one agent to reads takes a separate server in managed mode, or your client's tool settings.

Zapier MCP and alternatives

Zapier MCP vs n8n: Zapier gives agents ready-made actions across thousands of apps. n8n exposes workflows you build, and you can host it yourself. More hosted options are in what is an MCP gateway.

Zapier through Upsy

Upsy connects to Zapier MCP as one client with your Zapier sign-in, next to the apps you connect to Upsy directly. It's the same address as Upsy's other apps, https://upsy.ai/connect.

What it adds:

• Lookups without the write tool. With Pick actions, tick execute_zapier_read_action and the inspect and discover tools, and leave out execute_zapier_write_action. The agent can run search and read actions but not sends, updates or deletes. Keep "Find or Create" actions turned off in Zapier for that agent, since they can create records. Upsy's plain Read-only level doesn't run Zapier actions at all, because it treats any tool named "execute" as a change.
• A level per agent. One Zapier sign-in, with different tools for each agent, enforced by Upsy.
• Full tools where they exist. Connect Gmail, Slack, Notion, Linear and other apps to Upsy directly, and use Zapier for the rest.
• History and alerts: every Zapier tool an agent runs, listed under that agent.
• Any client: Claude, Claude Code, Cursor, Codex, ChatGPT, Grok, Muse, or anything with remote MCP and OAuth.

Every Zapier call through Upsy is still a Zapier tool call, at 2 tasks.

Connect Zapier to Upsy

1. Go to upsy.ai and click Set up Upsy.
2. Open Apps, click Connect an app and pick Zapier. Sign in to Zapier and allow access.
3. On the Assistants page, click Add assistant and choose your client under Which assistant? for its exact steps.

Add the server to your client

Signing in opens an Upsy page: tick Zapier and the agent's other apps, and click Allow. To allow reads only, use Pick actions on the agent's page afterwards.

• Claude Code: claude mcp add --transport http --scope user upsy https://upsy.ai/connect, then /mcp to sign in.
• Cursor: add {"mcpServers":{"upsy":{"url":"https://upsy.ai/connect"}}} to ~/.cursor/mcp.json, then click Connect next to upsy.
• Codex: codex mcp add upsy --url https://upsy.ai/connect, then codex mcp login upsy.
• Claude and ChatGPT: add https://upsy.ai/connect as a custom connector (Claude, ChatGPT).
• Grok and Muse: paste the setup prompt Upsy shows into a chat and approve the sign-in.

Server details, such as OAuth endpoints and token lifetimes, are in the Gmail MCP server guide; they're the same for Zapier.

Check that it works

Using Upsy, list the Zapier actions I have turned on.

The agent should list your enabled actions by app. The calls show in Upsy's History under that agent.

Prompts to try

• "Find the HubSpot contact for jane@acme.com and add a note about today's call."
• "Add these three expenses as rows in my Expenses sheet in Google Sheets."
• "Look up order 1042 in Shopify and tell me its shipping status."
• "Create a Trello card in the Launch board for each action item in this meeting summary."
• "Save a skill called weekly-report with the steps I just followed, so you can repeat them next week."

Troubleshooting

"Action execution was halted by the platform"

Your Zapier admin restricted that app or action, or the agent sent data Zapier couldn't use. Check your account's app restrictions, then try again with simpler input.

McpAuthorizationError

An app connection in Zapier needs signing in again. Reconnect that app in your Zapier account, then try again. With a token, generate a new one.

Tools return 404

The connection token was rotated or expired. Generate a new one under Settings › API token and update your client.

Sign-in finishes, but the client never connects

Start the sign-in from the client, not from Zapier's site, and update the client. Older Cursor and Windsurf versions, VS Code extensions, Claude Desktop and WSL are the usual causes.

The agent can't see an action

In agentic mode, ask it to search for and turn on the action. In managed mode, add it at mcp.zapier.com. Then refresh the client's tools: in ChatGPT, Refresh tools; in Claude Desktop, restart; in Slackbot, wait up to a day.

Calls fail after working for weeks

You may be out of tasks for the month. Check Billing & usage › Usage in Zapier, since MCP calls don't show in the Zap usage graph.

"Connection settings are owner-only"

The server was shared with you. Only its owner can connect clients or change settings, so use your own server.

"The user hasn't allowed you to use Zapier's …"

That's Upsy refusing a tool outside the agent's level, such as execute_zapier_write_action when you left it out. Change Zapier's level on the agent's page in Upsy.

Security and privacy

• Zapier holds your app logins. Agents call Zapier's tools; they never see the app credentials. Zapier is SOC 2 Type II certified and stores data in AWS US-East-1.
• History at mcp.zapier.com logs every call. Deleting a server deletes its logs.
• Your app permissions still apply, along with any app and action restrictions on your Zapier account.
• Lock fields that matter, like the recipient of an email or the channel for a message, so the agent can't change them.
• Tokens act for one server until you rotate them. Keep them out of shared config files.
• Prompt injection. An email or record an agent reads can carry instructions aimed at it, and Zapier's write tools can send data anywhere your connected apps reach. Give agents that read outside content no write tools.
• Through Upsy, agents never get your Zapier token: they call Upsy, and Upsy calls Zapier. Upsy emails you when one agent runs 100 or more reads in 10 minutes, and when a new agent or app connects. Its send and delete alerts go by tool names, so calls through execute_zapier_write_action don't count toward them.

FAQ

What is Zapier MCP?

A hosted MCP server from Zapier. AI agents connect to one address and run actions in 9,000+ apps through your Zapier account, without a separate integration for each app.

Is Zapier MCP free?

It's on every Zapier plan, Free included, but each tool call uses 2 tasks. Free has 100 tasks a month, shared with your Zaps, so about 50 calls.

How much does Zapier MCP cost?

Your Zapier plan's tasks pay for it: 2 per successful call, 0 for failed calls. There's no separate MCP price.

What's the Zapier MCP server URL?

https://mcp.zapier.com/api/v1/connect, for every client. Older per-server URLs and SSE addresses no longer apply.

Does Zapier MCP support SSE?

No. It uses Streamable HTTP only.

Can Zapier MCP run my existing Zaps?

Not directly. MCP runs actions when the agent calls them. With early access, agents can build and run Next Gen Zaps, and turn an existing Zap into one.

Can I make Zapier MCP read-only?

The closest is a managed-mode server with only search actions, and no "Find or Create" actions. You can also block execute_zapier_write_action in your client, or through Upsy pick execute_zapier_read_action and leave out the write tool.