Article

Limit Claude to chosen actions with Zapier MCP

For small-business teams using Claude, learn how to restrict Zapier actions, fix important inputs and budget the shared task allowance.

Editorial illustration for Limit Claude to chosen actions with Zapier MCP: a geometric block represents a model release. Not documentary evidence.

To limit Claude to chosen actions in business apps, use managed mode on its Zapier MCP server, remove unwanted actions and fix inputs that should not vary. Switching modes alone does not narrow the existing tool list: every enabled action remains available.

Zapier’s Claude tutorial, refreshed on November 20, 2026, highlights the appeal for small-business owners and operations teams: help with work such as preparing email drafts, while keeping sending under human control. This is a guide to an existing capability, not a new product launch. MCP, short for Model Context Protocol, is the connection standard that lets Claude call tools in other apps.

1. Find the Zapier server Claude actually uses

Zapier lists MCP on Free, Professional, Team and Enterprise. Its Claude connector guide also lists support for all Claude plans, including Free. Claude Team and Enterprise require an owner to enable Zapier first; each member must still sign in individually.

For a new connection, ask Claude to add the Zapier connector, select Connect and follow Zapier’s sign-in flow. The authorization screen includes copying a setup prompt before Continue. Review the requested access, select the intended Zapier account and grant access. For this fixed-action setup, pause before running the copied onboarding prompt: review the server’s scope first.

Sign-in connects a new client to your default Zapier server, creating one if necessary. At the Zapier MCP dashboard, inspect the server and its Clients list. If Claude is already connected, start there. Several of your AI clients can share the same server, so a change may affect more than Claude.

2. Switch to managed mode and review every action

New servers generally start in agentic mode, where the assistant can discover and enable actions. Managed mode instead supplies the actions you configure and does not let the assistant add new ones. That is the documented difference between the modes.

With Owner or Editor access, open the server’s Settings tab, select Managed mode and save the change. Then inspect the Apps tab and remove operations the job does not need. Zapier’s switching instructions say all previously enabled actions become individual tools; the switch is not an automatic cleanup.

The mode applies to every client on that server. Skills—saved workflow instructions—and code actions become unavailable, though they are retained for a switch back. Next Gen Zap workflow-building tools are also unavailable in managed mode. Choose this mode for a defined task, accepting that additional actions must be configured by a person.

3. Separate email drafting from sending

For a draft-only email workflow, allow lookup and draft creation, then remove every action that sends, forwards or deletes messages. In Zapier’s Gmail action catalogue, Find Email and Create Draft Reply are separate from Send Email and Forward Email.

  • Keep only the lookup and draft actions your particular job requires. Review the complete enabled list, not just the first few tools.

  • Exclude Find or Send Email from a lookup-only role. It is a search-or-write operation that can create a message when no match is found.

  • Check the account’s actual action names and required inputs. This is an illustrative configuration, not a promise that every reply needs exactly the same tools.

A draft-only workflow still writes data: it creates a draft. It also does not establish which messages a lookup can read. Permissions in the connected app and Zapier’s account restrictions still apply. As a practical consequence, removing a sending action from this server does not remove sending through another connector or another tool available to Claude.

For a related example of why allowed operations and reachable content are separate questions, see what Apify can access through a Notion connection.

4. Fix important inputs and check the account

As the server owner, open Apps, use the action’s three-dot menu and choose Configure. For inputs that must stay constant, choose the fixed-value option, enter the value and save. Zapier’s action configuration guide says that value is sent on every call, ignoring a different value supplied by the client.

For example, a customer-record action could always assign new leads to one owner. Leave genuinely variable content available to Claude. Where supported, restrict a dropdown to an approved set of choices; this option is limited to dropdown fields on write actions.

Also check the app’s default connection and any per-action override, so the action uses the intended work account. Field settings take effect without reconnecting and work in both modes. The practical distinction: removing an action controls what Claude can do; fixing an input constrains how an allowed action runs.

5. Refresh Claude and check the remaining permissions

After changing modes, restart Claude Desktop to refresh its cached tool list. Inspect the available tools before using them. If an expected managed-mode action is missing, check that it is enabled under Apps and that its app connection is authenticated. Zapier does not document an equivalent browser-only refresh sequence on that page.

Review Claude’s own tool approvals separately. Anthropic’s connector guidance allows disabling tools that are not needed. Do not assume every action will trigger a fresh confirmation: Research can invoke permitted connector tools without further approval, and Anthropic advises disabling write-capable tools for that mode. Draft creation is write-capable too.

Start with one low-impact item. Inspect the actual draft or record in the destination app, then check the server’s History tab. A recorded tool call is useful for review, but the business result still needs checking.

What does this use from your Zapier allowance?

As checked on November 20, an ordinary successful external-app call through MCP consumes two Zapier tasks, including successful test calls. Failed calls consume none. MCP shares the account’s allowance with other Zapier work; it has no separate subscription. The Free plan lists 100 tasks per month at $0.

Illustrative calculation: suppose 25 emails each require exactly one successful lookup and one successful draft-creation call. That is 25 × 2 calls × 2 tasks = 100 tasks. It would exhaust the Free allowance if nothing else used it. Extra successful lookups, retries or tests increase consumption; this is not a guarantee of 25 completed replies. Claude limits and third-party costs remain separate.

There are scheduled exceptions from December 1, 2026: free By Zapier apps such as Tables and Forms become unmetered through MCP, while AI by Zapier uses its own pricing. Those exceptions do not make ordinary Gmail or CRM calls free. For the wider change, see Zapier Next Gen and SDK billing on December 1.

Before recurring use, compare the MCP Tasks and Plan Tasks counters. At the allowance limit, calls stop unless the plan and pay-per-task settings permit capped, paid overages. Check these counters after the first limited run before deciding how many replies to process.