Zapier MCP: The Real Cost Is Tasks (Not Tokens)

Search this topic and you will read the same warning ten times: Zapier MCP connects your agent to 9,000+ apps and tens of thousands of actions, so watch out — all those tool definitions will devour your context window. That warning is describing a product that no longer exists by default. Zapier MCP now defaults to dynamic discovery and exposes 15 static meta-tools, in Zapier’s own words “regardless of which apps you have connected”. There is no 40,000-tool dump to fear. The cost is real, but it is metered somewhere else entirely — in tasks, at a fixed 2 per successful call, which is exactly double what the same action costs inside an ordinary Zap. Here is the version built from Zapier’s primary docs, as of July 2026.

What Zapier MCP actually is

The Model Context Protocol is an open protocol for connecting AI assistants to external systems. One correction before anything else, because it is everywhere: this is not Anthropic’s protocol any more. MCP was donated to the Linux Foundation’s Agentic AI Foundation on 9 December 2025 — a directed fund co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg, where MCP is a founding project alongside Block’s goose and OpenAI’s AGENTS.md.

The opposite over-correction is also wrong. Nothing about day-to-day governance changed: per MCP’s own announcement, “The governance model we introduced earlier this year continues as is”, the Linux Foundation “will not dictate the technical direction of MCP”, and existing maintainers keep decision-making authority through the SEP process. Anthropic continues to fund it and maintain core infrastructure. So: not Anthropic-owned, not Anthropic-abandoned. The current protocol revision is 2025-11-25 — check /specification/versioning rather than any blog post, because the older 2025-06-18 page still resolves and still looks authoritative, which is exactly how stale version numbers propagate.

Zapier MCP, then, is a hosted remote MCP server that puts Zapier’s integration catalogue behind that protocol. Zapier’s docs describe the reach as “9,000+ apps and 40,000+ actions”. Worth noting honestly: Zapier’s own marketing page says 30,000+ actions instead. Two primary sources from the same vendor disagree, so treat the count as tens of thousands and move on — the number is not load-bearing for any decision you will make. If you want the wider argument about when a protocol beats just calling the endpoint, we covered it in MCP vs API.

The 15-tool correction

Here is the claim that most published writing on this keyword gets wrong, and it inverts the whole article.

Zapier MCP defaults to dynamic tool discovery. Zapier’s docs are unambiguous: “Your server exposes 15 static meta-tools across five categories. These are always available regardless of which apps you have connected.” We counted the rows in Zapier’s own table — there are exactly 15, so the stated count is self-consistent. They break down as:

• Action management (6) — list_enabled_zapier_actions, discover_zapier_actions, enable_zapier_action, disable_zapier_action, auto_provision_mcp, write_code_action.
• Execution (2) — execute_zapier_read_action, execute_zapier_write_action.
• Configuration (1) — get_configuration_url.
• Skills (5) — list_zapier_skills, get_zapier_skill, create_zapier_skill, update_zapier_skill, delete_zapier_skill.
• Feedback (1) — send_feedback.

Six plus two plus one plus five plus one is fifteen. Connect one app or connect four hundred; the surface your model sees at session start is the same fifteen. The agent finds what it needs at runtime by calling discover_zapier_actions, turning it on with enable_zapier_action, and firing it with execute_zapier_write_action. That is a search index, not a catalogue dump.

Why does every article say otherwise? Because they are describing an older Zapier MCP, where you hand-picked actions in a web UI and each one became a dedicated tool. That mode still exists — Zapier calls it manual configuration and it is genuinely useful, which we get to below — but it is not what you get when you connect today. When you authenticate over OAuth, Zapier says auto_provision_mcp runs automatically and configures the server from apps you have already connected. One caveat that will bite a team: it covers only your own connections. Apps shared with you by another Zapier user are not included.

The fifth category is worth a line of its own because nobody mentions it. Zapier Skills are reusable, Markdown-based workflow instructions your agent can load on demand, managed through those five meta-tools or the Skills tab at mcp.zapier.com, and Zapier ships packaged ones including an onboarding Skill. It is a prompt-library primitive bolted onto a tool server, and it is the part of the product with no obvious analogue in the rest of the MCP server ecosystem.

The 2-task tax

So if the tool surface is not the cost, what is? Zapier answers this itself, in a sentence buried in a usage doc that nobody quotes:

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

That single rate is the entire cost model. It is confirmed in four separate Zapier primary sources — the MCP usage doc, the general help-centre article on task measurement, the MCP security doc, and the zapier.com/mcp marketing page, which adds that it is “the same task bucket your Zaps use. No separate SKU, no new procurement.” And the docs make the point that nothing else is metered at all: “There is no limit on the number of tools you can add to a server, or the number of tool calls you can make per session. The only usage limit is the task allowance included with your Zapier plan.”

Now put that next to the other rate Zapier publishes, in a different document: an action step inside a Zap counts as one task.

The derivation: 1 task per Zap action step; 2 tasks per MCP tool call. Therefore performing the same action through MCP costs exactly 2x what it costs inside a Zap. Zapier never states this comparison anywhere. It is arithmetic on Zapier’s own two published rates — both sources are theirs, the sentence is ours.

That is not a gotcha. It is a price for a service: discovery, credential storage, refresh, and the right to not know which app you will need until the model decides. But you should know you are paying it, because nobody tells you.

What 100 tasks actually buys

Zapier’s free plan includes 100 tasks per month — confirmed on the pricing page, and the one allowance we will print as fact. At 2 tasks per successful call, that is a hard ceiling of 50 successful MCP tool calls per month. Not 50 conversations. Fifty calls.

Zapier’s docs supply their own worked examples of agentic loops, and they are more damning than anything a critic would write. We have added the third column:

The first three rows are Zapier’s figures verbatim; the percentages are ours. One instruction — search and update 10 records — burns 22% of a free month. You get roughly four of those and the account stops responding until the allowance resets.

Give Zapier credit where it is due, though, because it changes how you should design the loop: failed tool calls do not consume tasks. An agent that gropes around, mis-shapes a payload, retries and eventually gets it right pays only for the successes. Exploration is genuinely cheaper than the meter suggests. Successful fan-out is not. That asymmetry is the single most useful thing in Zapier’s usage doc and the reason a chatty, self-correcting agent may cost less than you fear while a tidy, competent one that touches 40 records costs more than you budgeted.

Two more mechanics worth knowing before you plan around this. Test tool calls count. Tasks consumed by a server you later delete still count for that billing period. And at the ceiling, Zapier says MCP tool calls simply stop working until the allowance resets or you upgrade — server config and connected apps persist, so there is nothing to rebuild.

On paid tiers we are going to be deliberately vague, and you should be suspicious of anyone who is not. Zapier’s pricing page renders task allowances through a JavaScript slider, which means the confident per-tier task tables circulating in third-party posts were not read from source. What we can confirm from the page itself: pricing is task-based, tiers span 100 per month up to 2M per month with a custom option above, Professional starts at $19.99/mo, Team starts at $69/mo, a 33% annual discount is advertised, and Enterprise is custom. Divide whichever allowance you are actually shown by two. That is your call budget.

Where the token cost is real

The tool-bloat warning is not invented. It is misfiled. There are exactly two places it still holds, and naming them precisely is the whole point.

One: manual configuration mode. Zapier’s docs: “Some servers use manual configuration instead of dynamic discovery. In this mode, each action is exposed as a dedicated tool.” That is where the surface people fear actually materialises — and it exists only because you opted into it. Zapier lists real benefits for it: a fixed, predictable toolset, tightly scoped access where the agent cannot enable new actions on its own, and pre-configured or locked field values. Those are good reasons. What Zapier’s docs say nothing about is the token implication, so say it here: manual mode buys determinism and pays for it in context. That is a defensible trade if you are shipping a narrow production agent and a bad one if you are exploring.

Two: gateway and proxy users. Which needs its own section, because it is us.

The gateway exception (yes, that includes us)

Anthropic’s Claude Code documentation describes a feature called tool search, and it is why the “MCP eats your context” claim is false under defaults: “Tool search keeps MCP context usage low by deferring tool definitions until Claude needs them. Only tool names and server instructions load at session start, so adding more MCP servers has minimal impact on your context window.” It is on by default.

Then, from the same page, the sentence that concerns our readers directly:

“It is also disabled when ANTHROPIC_BASE_URL points to a non-first-party host, since most proxies don’t forward tool_reference blocks.”

DataLLM Lab is a non-first-party host. We are an LLM gateway. If you run Claude Code through our endpoint — which is the entire point of us — tool search switches off and you pay the full, undeferred cost of every tool definition at session start. The warning you read elsewhere is wrong for most people and right for our own users. Writing that softly would be worse than not writing it.

The constructive half is that it makes Zapier’s default the correct choice for exactly this cohort. Fifteen meta-tools load upfront without drama. A hundred manually configured Gmail-and-Sheets tools would not. If you are behind any gateway, stay in dynamic mode — you are the reader for whom that advice has teeth.

Some accuracy discipline on this, because it is the most volatile fact in the article. Tool search is Claude Code behaviour, not an MCP spec feature; it rides on tool_reference blocks, a beta-gated API capability requiring Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 or later. It is also off by default on Google Cloud’s Agent Platform. You can force it with ENABLE_TOOL_SEARCH, except that CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS keeps it off and ENABLE_TOOL_SEARCH cannot override that. ENABLE_TOOL_SEARCH=auto loads schemas upfront when they fit within 10% of the context window and defers only the overflow. Claude Code imposes no fixed per-server tool cap; the practical limit is your context budget. All of that is env-var-and-beta-header territory, dated July 2026 — check the docs, not us, before you build policy on it.

One adjacent limit that bites Zapier specifically: Claude Code warns when MCP tool output exceeds 10,000 tokens and caps output at 25,000 tokens by default, tunable via MAX_MCP_OUTPUT_TOKENS. A Zapier read action against a busy Gmail account will meet that ceiling long before it meets a task ceiling. The general playbook in how to cut LLM API costs applies unchanged: the cheapest token is the one you never put in the window.

Transport and auth gotchas

Three things here will cost you an afternoon if you learn them the hard way.

Streamable HTTP or nothing. Zapier’s client docs say Zapier MCP works with any client that supports MCP over Streamable HTTP, then say it flatly: “Zapier MCP does not support SSE.” No stdio either, since the server is remote. For context, the current 2025-11-25 spec defines exactly two standard transports — stdio and Streamable HTTP — and notes Streamable HTTP “replaces the HTTP+SSE transport from protocol version 2024-11-05”. Zapier has simply declined to support the deprecated one. The documented client list covers Claude, Claude Code, Claude Cowork, ChatGPT, Cursor, VS Code and Windsurf, connected from inside the client with, in Zapier’s words, no server setup required; most connect through OAuth.

Three auth methods, for three different situations. A client-specific server URL generated during setup, OAuth-based, for the listed clients. A connection token for unlisted or custom clients — pick “Other” when creating the server. An API key, and this one is narrower than it sounds: Zapier says it is for TypeScript and Python SDK integrations only and is “not used for standard AI client connections.” For the token and API-key paths the endpoint is exactly https://mcp.zapier.com/api/v1/connect with an Authorization: Bearer header. Tokens are shown once and cannot be retrieved later; rotating one immediately invalidates the old one and breaks any client still holding it.

Use the header, not the URL. Zapier documents two ways to pass the connection token and marks the Authorization header as “Option 1 (Recommended)”. Option 2 puts the token in a URL query parameter. Take the recommendation seriously, because MCP’s own authorization spec says access tokens “MUST NOT be included in the URI query string”. We are not going to claim Zapier violates the spec — a connection-token query parameter is a Zapier connection convenience, not necessarily an OAuth access token flow, and Zapier already steers you to the header. But the tension is real, query strings leak into logs and referrers, and Zapier has told you which one it prefers. Take it.

While we are in the spec: MCP authorization is optional, defined for HTTP transports only, and stdio implementations should not use it. It is built on OAuth 2.1, with PKCE using S256 as a MUST and RFC 8707 resource indicators as a MUST, and token passthrough explicitly forbidden. One drift worth knowing if you last read this a year ago: Dynamic Client Registration (RFC 7591) is now only a MAY, retained for backwards compatibility, with OAuth Client ID Metadata Documents as the preferred path for parties with no prior relationship.

Governance default, stated precisely. Zapier MCP is enabled by default for all accounts, including Enterprise, and access is granted at the account level, so every user in the account can create and configure their own servers. Disabling it at workspace level requires contacting an account manager. Account-wide app and action restrictions are enforced in MCP, but per the docs they cannot currently be set exclusively for Zapier MCP — any restriction applies across all Zapier features. That is an access-control default, not a vulnerability, but if your compliance story assumes MCP is opt-in, it is not. On the reassuring side: data sits in AWS US-East-1 (no region-specific residency unless agreed, no dedicated VPC, no on-prem), SOC 2 Type II is documented via trust.zapier.com, Enterprise is opted out of AI training by default, and the tool-poisoning surface is closed by construction — “Users cannot bring tools in from third-party sources. All tools are owned and controlled by Zapier.” The History tab at mcp.zapier.com gives per-call activity logs, MCP events land in the account audit log, and superadmins can review any user’s logs.

Zapier MCP vs n8n vs your own server

The comparison that matters is not feature-for-feature. It is what unit each one meters, because that is what decides your bill at volume. Zapier bills per action, at double rate for MCP. n8n bills per workflow execution regardless of how many steps are inside it. Your own server bills nothing and costs you an on-call rotation.

A licensing note, because this one is mangled constantly: n8n is not open source and does not claim to be. It is under the Sustainable Use License, a fair-code license n8n wrote in 2022, which limits use and modification to internal business purposes or non-commercial and personal use, and requires that redistribution be free and non-commercial. OSI open source licenses cannot restrict use, which is precisely why n8n declines the label. Files with .ee. in the filename or .ee in the directory name require an n8n Enterprise License. “Self-host it for free” is true of the Community Edition and only of the Community Edition.

On price we are going to refuse the apples-to-apples table you want. n8n publishes EUR figures — Starter at €20/mo for 2.5K executions and Pro at €50/mo for 10K, billed annually, with unlimited users, workflows and integrations on all plans — and Zapier publishes USD figures against a completely different unit. A workflow that touches 12 records is one n8n execution and 24 Zapier tasks. Comparing €50 to $19.99 tells you nothing; comparing units tells you everything. That structural point, not the numbers, is the reason to read our n8n MCP write-up alongside this one.

And an interop trap worth flagging before you wire them together: Zapier MCP is Streamable-HTTP-only and rejects SSE, while n8n’s MCP Client Tool node docs still describe only an SSE Endpoint parameter. Streamable HTTP was added to that node by a merged PR (#15454), keeping deprecated SSE for compatibility — so n8n-to-Zapier works, but only on a recent enough n8n, and the docs page lags the code. We are not printing a version number, because the ones circulating are community-sourced. Open the node in your instance and look at what the UI actually offers.

The decision rule

Count two numbers: actions per run, and runs per month. Multiply, double it, and compare to your allowance. That is the whole analysis.

Zapier MCP wins when the work is broad, occasional and long-tail. You need one of thousands of apps and you do not know which until the model decides. Somebody else owning OAuth refresh for a niche CRM is worth real money. Volume is low enough that 2 tasks per call is noise. This is the case Zapier MCP was built for and it is a good product for it — the discovery loop plus stored credentials is genuinely hard to rebuild.

It loses precisely where automation usually ends up: high-volume, multi-step, predictable. Per-action metering at double rate compounds, and every extra record the agent touches is another 2 tasks. A single n8n execution is metered once no matter how many steps it contains. A direct API call is metered zero.

The rule: if you can name the API in advance and you will call it more than a few hundred times a month, Zapier MCP is the wrong layer. Its value is discovery and auth, not throughput. The corollary: if you are behind a gateway, stay in dynamic mode (15 tools survive the loss of tool search); if you need deterministic scoping and can afford the definitions, manual mode buys predictability at a context cost.

Which is a specific instance of a general truth about agent architecture: keep the deterministic parts deterministic and spend the expensive, metered, probabilistic layer only where judgement is actually required. The same instinct decides which model you point at an agent loop. A model that fumbles three calls before succeeding costs you nothing extra on Zapier’s meter — failures are free — but it costs you tokens, latency and patience. A model that gets it right first time and then cheerfully fans out across 40 records costs you 80 tasks.

Last thing, and it is the reason to date every claim in here: the 2-task rate is a pricing term. It is quadruple-confirmed across Zapier’s docs today and described as fixed, which is about as solid as a vendor rate gets. It is also exactly the kind of number that changes quietly and gets inherited by every article that copied it. Ours says July 2026 on it. Check docs.zapier.com/mcp/usage before you build a budget on our arithmetic.

FAQ