How much does it cost to give an AI agent access to your automation platform?

What one agent tool call costs

Verified against vendor documentation on 26 August 2026.

Zapier is the outlier in both directions. It is the only platform that charges a documented surcharge for the privilege of an agent making the call, and the only one that states plainly that failures are free.

Zapier: the only explicit per-call rate

Zapier's MCP usage documentation is unusually direct. There is no separate MCP subscription; the server "runs on your existing Zapier plan", drawing from the same task allowance as Zap workflows. Against that pool:

• Each successful tool call consumes two tasks, at a fixed rate.
• Tool calls that fail consume nothing.
• There is no limit on the number of tool calls per session, and no limit on the number of tools added to a server.

The arithmetic that follows is the part worth internalising. Zapier's free plan carries 100 tasks per month, which at two tasks per call is 50 successful agent tool calls in a month. An agent exploring a schema, listing records and then writing one row has spent three calls and six tasks before it has done anything useful. Teams evaluating an agent against Zapier on a free plan routinely exhaust the allowance during setup and read the resulting failures as a broken integration.

For current plan pricing and allowances above the free tier, see Zapier's official pricing page.

Where the meter is the work, not the call

The other three platforms in this comparison bill the consequence rather than the request.

Make replaced operations with credits as its billing unit, and its help documentation states that for typical modules "1 operation equals 1 credit". Some modules cost more: Make notes fixed higher credit usage on certain AI content extractor modules, and dynamic credit usage on some AI and advanced apps that varies per run. A scenario an agent triggers over MCP therefore costs whatever that scenario costs, which is a function of module count rather than call count. Make's documentation surveyed on 26 August 2026 states no separate MCP charge.

n8n meters cloud plans by workflow execution. An MCP call that runs a workflow produces a workflow execution, which is the unit that counts; n8n publishes no MCP-specific surcharge.

Pipedream confirms that MCP tool calls consume credits, with credits defined as "one credit per 30 seconds of compute time at 256MB megabytes of memory". A fast tool call and a slow one are not the same price, and doubling a workflow's memory doubles credit consumption for the same duration.

The practical difference: on Zapier the cost of an agent scales with how chatty it is, and on Make, n8n and Pipedream it scales with how much work it asks for. A verbose agent making many small reads is expensive on Zapier and cheap elsewhere. An agent triggering one heavy twenty-module scenario is the reverse.

Why agent retries change the arithmetic

Coding agents retry. A tool call that returns a malformed argument error is usually re-issued with corrected arguments, and an agent working through an unfamiliar schema may make several attempts before the first success.

Zapier's rule that failed calls are free is what makes this survivable on that platform, and it is the single most useful line in its usage documentation.

Make carries the opposite hazard, and it is a timeout rather than a billing rule. Make's developer documentation puts the scenario-run tool timeout at 25 seconds over OAuth and 40 seconds with an MCP token, while noting that the called scenario "continues running in Make for up to 40 minutes" after that timeout, retrievable by executionId. An agent that treats the timeout as a failure and retries has started a second scenario run while the first is still executing. Both consume credits. Nothing in the agent's transcript will show it.

The mitigation is to have the agent retrieve the execution by executionId rather than re-issue the call, which is the reason the field is returned.

Paying with a plan tier rather than a unit

Some of the cost is access, not consumption.

What the vendors do not document

Three gaps are worth naming, because their absence is itself decision-relevant as of 26 August 2026.

1. Per-call rates outside Zapier. No other platform in this comparison publishes a documented MCP surcharge. That is consistent with billing the underlying work, and it is not the same as a published guarantee that no surcharge exists.
2. Failure handling outside Zapier. Only Zapier states what a failed tool call costs.
3. Rate limits. Zapier's usage documentation mentions none, and states there is no cap on calls per session or tools per server.

Anyone budgeting an agent against a platform other than Zapier should meter a representative week rather than extrapolate from a published rate, because the published rate does not exist.

Editor's Note: The number that changes how teams plan is Zapier's two tasks per successful call. Working it against Zapier's own free-plan allowance of 100 tasks a month on 26 August 2026 gives 50 successful agent tool calls, and an agent that reads a schema before it writes anything can spend a third of that in one exchange. We now budget agent access as a separate line from the automations themselves rather than assuming an existing plan absorbs it. The caveat: this arithmetic is drawn from vendor documentation, not from a metered production run, and an agent's real call volume depends on how much context it is given before it starts. Meter a week before committing to a tier. — Rafal Fila, ShadowGen