Zapier MCP: every AI tool call costs two tasks, not one

Bottom line. Zapier’s MCP documentation is explicit: each successful AI tool call through an MCP server consumes two tasks, at a fixed rate, with no per-session cap. On a Professional plan starting at $19.99 a month for 750 tasks, that is 375 agent actions. Wire Claude or ChatGPT into Zapier without doubling your call-count math and the plan runs dry mid-month.

Zapier’s MCP server is the bridge most small operators use to let an AI assistant act in their other tools: write the lead to the CRM, file the invoice, update the sheet. The billing rule that governs it lives in a help-centre page updated 4 Jun 2026, and it is unambiguous:

“Each successful tool call through your MCP server consumes two tasks. This is a fixed rate.”

— Zapier’s MCP usage documentation, updated 4 Jun 2026.

There is no separate MCP billing, no volume discount on the multiplier, and no cap on how many tool calls an AI session can make. Failed calls do not bill. Successful ones always do, whether the action was a one-cell update or a full record write.

Rates from Zapier’s documentation and pricing page; the capacity rows are arithmetic on those rates.

The doubling is the budget

The trap is not the price, it is the invisible multiplier. An operator who budgets “750 tasks, my agent does about 20 things a day, easily fits” is budgeting at half the real burn: 20 successful calls is 40 tasks, and a month of that is 1,200 tasks against a 750-task plan. The workflow did not change; the unit of account did.

Agents also make more calls than their flowchart suggests. A single “qualify this lead and add it to the CRM” run can be a look-up, an enrichment call, a duplicate check and a write, four successful calls, eight tasks. This is the same meter-shock pattern we walked through when Notion’s agents went metered and in the broader bill-rising read: the per-unit price looks trivial until an autonomous system multiplies the units.

Where the math still works

Two tasks is a few cents on most plans, and a few cents for a lead written into the CRM or an invoice filed is an excellent trade. The MCP path earns its rate on actions. Where it breaks is chatter: an agent polling for status, re-syncing data it could batch, or confirming things nobody asked it to confirm. That work belongs in ordinary scheduled Zaps, where a task is a task, or in a batched job, the discipline the bootstrapped-SaaS cost read applies stack-wide. If you are building the agent side from scratch, the no-code agent-building read covers the build half, and our sibling piece on Notion Workers’ free window covers the cheapest current alternative for Notion-centric syncs.

One unhedged line: before any Zapier-connected agent goes live, divide your task allowance by two and write that number down as the real capacity. Budget against 375, not 750, and meter the first week, because the observed call count, not the flowchart, is the bill.

What changes this verdict

Cadence on this piece is 30 days, because automation-platform billing rules are moving fast across the board. Three changes would move it: Zapier changing the two-task rate, capping or itemising MCP usage separately; Zapier adding an MCP-specific plan or pool that decouples agent calls from the task allowance; or a competing automation platform pricing MCP calls at parity with ordinary tasks, which would change the comparison for new builds. We re-test against the documentation on or before 9 Jul 2026; the Holding-up record for OPS-098 carries any change, dated.