Zapier's hosted MCP server. The one endpoint, the two authentication paths, the meta-tools that let an agent discover and enable actions on the fly, and what each call costs, read from docs.zapier.com on 2 September 2026.
Interactive examples · No account is connected on this page
Updated
Zapier MCP is a hosted Model Context Protocol server that gives an AI client access to Zapier's app integrations. Every client connects to the same Streamable HTTP endpoint, https://mcp.zapier.com/api/v1/connect. Clients on Zapier's supported list authenticate with OAuth and Zapier provisions the server during sign-in; anything else uses a connection token. Each successful tool call consumes two tasks from your Zapier plan.
Pick an example to see which Zapier MCP tools a task would call and what the result looks like. Examples are illustrative; nothing on this page connects to Zapier.
Zapier MCP is hosted at https://mcp.zapier.com/api/v1/connect over Streamable HTTP. For supported clients, connect from inside the client and sign in, and Zapier creates the server for you. Claude Code takes one command; Cursor uses the Zapier plugin. Unlisted clients and your own code use a connection token generated at mcp.zapier.com.
The documented default is OAuth started in the AI client. Zapier creates and configures the server for you during sign-in and names it after the client, so the docs specifically tell you not to go and create a server at mcp.zapier.com first. Zapier's supported list includes Claude, Claude Code, ChatGPT, Cursor, VS Code, Gemini CLI, Windsurf, Zed, Warp, Replit, Kiro, Linear, Mistral, Microsoft Copilot Studio and more.
https://mcp.zapier.com/api/v1/connectClaude Code connects straight through the CLI over HTTP transport, then you restart it and authenticate with your Zapier account. Zapier's own setup page notes that on a Claude Team or Enterprise plan your organisation may require Owner approval before you can add MCP servers at all.
claude mcp add --transport http "Zapier-MCP" https://mcp.zapier.com/api/v1/connectFor Cursor the documented route is the Zapier plugin from the Cursor marketplace, then a prompt pasted into an Agent chat that runs the onboarding skill. Zapier warns that some Cursor versions carry an RFC 9728 OAuth issue that stops authentication completing, and the fix is to update Cursor and retry.
Set up Zapier MCP for me. Call get_zapier_skill with name "zapier:onboarding" and follow its instructions. If authentication is needed, help me through it, then retry.This is the one path where you do create the server first: go to mcp.zapier.com, add a server, choose Other, and generate a token from the Connect tab. The token goes to the same endpoint, preferably as an Authorization header rather than a query parameter, because a URL is more likely to end up in a log, a shell history or a committed config file. It is shown once, it is long-lived, and regenerating it invalidates the old one immediately.
Authorization: Bearer YOUR_CONNECTION_TOKEN
https://mcp.zapier.com/api/v1/connect?token=YOUR_CONNECTION_TOKENWhen you connect over OAuth, the auto_provision_mcp tool runs automatically and sets the server up from the apps already connected in your Zapier account. It covers your own connections only — apps shared with you by another Zapier user are not included. If your account has no connected apps the server starts empty, and the agent connects an app the first time it needs one.
auto_provision_mcpIn the default dynamic-discovery mode the agent manages its own toolset during the conversation. It searches for apps and actions, enables a specific one, then runs it — reads and writes go through separate execution tools, which is the seam to use if you want to reason about what an agent is allowed to do.
discover_zapier_actions
enable_zapier_action
execute_zapier_read_action
execute_zapier_write_actionOfficial documentation: Read the Zapier MCP docs.
Zapier MCP authenticates in two ways against the same endpoint. OAuth from the client suits anything on Zapier's supported list: Zapier creates the server during sign-in and provisions tools from your connected apps. A connection token suits unlisted clients and your own code: you create the server yourself, and the long-lived token is shown once.
| What differs | OAuth from the client | Connection token |
|---|---|---|
| Who it is for | Any client on Zapier's supported list | A client that is not on the list, or your own Python or TypeScript code |
| Who creates the server | Zapier, during sign-in | You, at mcp.zapier.com |
| Endpoint | https://mcp.zapier.com/api/v1/connect | https://mcp.zapier.com/api/v1/connect |
| Credential you manage | None; the client stores and refreshes the token | A long-lived token, shown once, tied to that one server |
| Initial toolset | auto_provision_mcp runs automatically from your connected apps | You configure it, or the agent discovers actions itself |
| How to revoke | Disconnect from the client or from mcp.zapier.com | Regenerate the token, which invalidates the previous one immediately |
Endpoint, transport, auth paths and meta-tool names transcribed from docs.zapier.com, read 2026-09-02. Zapier supports Streamable HTTP only; a client that can use SSE alone cannot connect.
Zapier MCP exposes the same 14 meta-tools whatever apps you connect. In the default dynamic-discovery mode the agent searches for, enables and runs actions itself, with reads and writes split into separate tools. Skills store reusable instructions, manual configuration fixes the toolset instead, and each named client gets its own server.
Whatever apps you have connected, the server exposes the same 14 meta-tools across five categories: action management, execution, configuration, skills and feedback. They are the fixed surface an agent sees, and everything app-specific is reached through them.
Most servers run in dynamic discovery mode, so the agent adapts mid-conversation rather than needing a pre-configured tool list. inspect_zapier_actions lists what is enabled, discover_zapier_actions searches for more, enable_zapier_action turns one on and disable_zapier_action removes it.
execute_zapier_read_action runs a search or lookup — find an email, look up a contact. execute_zapier_write_action runs anything that changes state — send a message, create a task, update a record. Two tools, so read-only intent is expressible in the call itself.
Skills are markdown workflow instructions the agent loads on demand, managed with list_zapier_skills, get_zapier_skill, create_zapier_skill, update_zapier_skill and delete_zapier_skill, or in the Skills tab at mcp.zapier.com. Zapier ships packaged skills including an onboarding one.
Some servers use manual configuration instead. Each action becomes a dedicated tool, configured at mcp.zapier.com rather than through the agent. The documented reasons to choose it: a predictable toolset, tightly scoped access with no ability for the agent to enable new actions, and pre-configured field values locked per tool.
Each AI client gets its own MCP server — one for Cursor, one for Claude, one for ChatGPT. You can run several, but only one per named client. There is no limit on actions per server or tool calls per session; the only ceiling is your plan's task allowance.
Zapier MCP is a hosted Model Context Protocol server that gives an AI client access to Zapier’s app integrations. Every client connects to the same Streamable HTTP endpoint, https://mcp.zapier.com/api/v1/connect.
Instead of a fixed tool per app, the server exposes 14 meta-tools that discover, enable and run actions. Each successful tool call uses two tasks from your Zapier plan; failed calls do not count.
Zapier MCP is the broad connector, reaching thousands of apps through one endpoint. The guides below cover the single-purpose servers an agent usually pairs it with when it needs deeper access to one system: n8n, Slack, Notion, GitHub, Cursor and Playwright. Each guide lists install steps, authentication and exposed tools.
Slack MCP ServerSlack's own server for channels, threads and search.
Notion MCP ServerNotion's hosted server for pages and databases.
GitHub MCP ServerRepositories, issues, pull requests and CI logs.
Cursor MCPHow Cursor loads MCP servers from mcp.json.Playwright MCP ServerBrowser automation for sites with no API.Supported clients start OAuth from inside the client and Zapier provisions the server during sign-in.
Unlisted clients and your own code use a connection token in the Authorization header.
Run one lookup before enabling write actions, and watch task usage.
Discover and enable actions for the apps you already connected in Zapier, then run reads and writes as separate calls.
Use manual configuration so each action is a dedicated tool with pre-set field values and no ability for the agent to enable more.
Store markdown workflow instructions as Zapier skills and let the agent load them on demand.
Combine Zapier with n8n MCP or Slack MCP depending on where the workflow lives.
These answers cover the practical questions about running Zapier MCP: what it costs, whether you must create a server at mcp.zapier.com first, which transport it uses, how to stop an agent enabling new actions, whether a server has an action limit, and where to see what the agent actually did.
There is no separate MCP billing — it runs on your existing Zapier plan. Each successful tool call consumes two tasks, and Zapier calls that a fixed rate. Failed calls consume nothing, and asking what tools are available, authenticating, or viewing history does not count. Batch work multiplies accordingly: Zapier's own example is that adding five rows to a spreadsheet is five tool calls, or ten tasks. When you hit the plan's task limit, MCP tool calls stop working until the allowance resets or you upgrade, but the server configuration stays in place.
Usually not, and the docs are firm about it. For any client on the supported list the flow starts inside the client: you add Zapier as a connector, sign in, and Zapier creates and configures the server during that sign-in. You create the server yourself only on the connection-token path — an unlisted client, or your own Python or TypeScript code.
Model Context Protocol over Streamable HTTP, and only that. A client that can use SSE alone cannot connect. If your client offers a transport choice, pick Streamable HTTP; if it was configured against an older SSE endpoint, repoint it at https://mcp.zapier.com/api/v1/connect.
Switch that server to manual configuration. In that mode each action is exposed as a dedicated tool and you configure the tools at mcp.zapier.com rather than through the agent, which is what the docs recommend when you need a fixed, predictable toolset, tightly scoped access with no agent-side enabling, or pre-configured field values locked per tool.
No. Zapier's documentation says a server has no limit on actions or tool calls, and that the only usage limit is the task allowance on your plan. The structural limit is elsewhere: one MCP server per named client, so a Cursor server and a Claude server are separate servers with separate configurations.
Every tool call runs through your existing Zapier app connections, so each one lands in the History tab at mcp.zapier.com alongside the rest of your Zapier activity. The sidebar there also shows Plan Tasks and MCP Tasks, which separates total account usage from what MCP specifically consumed. Deleting a server deletes its logs, though the tasks it already used still count for the billing period.
Dynamic discovery solves the wiring. What it does not decide is which steps should run without a human, on what trigger, and what happens when one of them fails. Those choices matter more with Zapier than with a single-purpose server, because each action an agent takes can consume tasks from your Zapier plan.