MCP Connector Costs: The Four Patterns Behind Every Surprise Bill

Almost every MCP connector is free. That sentence is true, and it is also the reason people end up with a bill they did not plan for.

The connector is a doorway. What costs money is the room behind it — the plan of the product it reaches into, the metered operations it quietly triggers, or the trial capacity it was built on before that capacity shrank. Anthropic says Claude’s connector directory now lists more than 950 MCP servers, so checking them one at a time is not a strategy. Reading the pattern is.

Every connector in our MCP directory has a setup guide behind it, and three of them are on our card as we write this: Canva Pro, Supermetrics and Windsor.ai. There are only four ways an MCP connector can cost you money. Learn to spot which one you are looking at, and the bill stops being a surprise.

The four cost patterns

Pattern 1 — Free, and it stays free

The connector rides on a product that is genuinely free for one person. Gmail, Google Drive, Google Calendar, Slack, Notion. You connect it, you use it, and nothing ever happens to your card.

The risk here is not financial, and it is worth naming anyway: these are the connectors that touch your most personal data, and several of them can write, not just read. Free and safe are different questions. This chapter answers the first one.

Pattern 2 — Free tier with a cliff

The connector is generous on day one and much smaller on day thirty-one. This is the pattern that catches people, because nothing breaks at signup. It breaks weeks later, once you have already built something on capacity you no longer have.

Windsor.ai is the clearest example we have tested. The free trial gives you ten data sources for thirty days. The permanent free plan that follows gives you one. We had four sources connected — GA4, Instagram, Pinterest and Meta Ads — when the account flipped to Free overnight on 10 August 2026.

Nothing disconnected. All four sources were still listed as in use, and every query we sent still succeeded — no error, no exception, no warning field. What changed was the answer. The numbers came back as zero. The date range we asked for was replaced with a single row dated today, no matter which of three different ways we asked for a month of history. And where a traffic source or a campaign name should have appeared, the text field instead contained an advertisement: “Uh-oh! You’ve connected more accounts than your Free plan allows. Upgrade here…” Ask an assistant reading that payload where your traffic came from, and it will tell you, in a perfectly grammatical sentence, that your top source is an upgrade prompt that sent you zero visitors.

It’s a threshold, not a proportion: three sources against a one-source plan behaved identically to four. At exactly one source, the same query returned thirty days of real data, correctly labelled. The free plan works fine inside its fence — it just never tells you when you’ve stepped outside it.

That is the choice Pattern 2 forces, and it is worth seeing clearly: pay, or keep answers you cannot trust. We paid — Windsor Basic, the day the cliff fired — because the alternative was publishing numbers a connector was quietly inventing.

The failure mode of Pattern 2 is not an invoice, and it isn’t quite a broken workflow either — a broken workflow announces itself. It’s a working workflow with wrong answers inside it, which is a harder thing to catch and a worse one to have missed.

Pattern 3 — Free connector, paid host

The MCP costs nothing. The plan you need in order to have anything for it to read costs plenty.

Supermetrics is the honest version: the connector is free, and the Starter plan we took in order to keep using it runs $55 a month, or $44 a month if you commit annually — €59.78 with VAT on our own invoice. That price buys one core destination, chosen from a list that includes Looker Studio, Sheets, Excel, Power BI, Claude, ChatGPT and Copilot. Ours wasn’t Claude, so the connector answers metadata calls fine and refuses every data call with the same error: DESTINATION_NOT_ON_LICENSE. A second destination costs another $62 a month. What the plan does include: three data sources, three accounts per source, one user, and up to 50,000 rows a month through the Data API and MCP access. Canva is the softer version: the connector is free and works fine on the free plan, right up to the monthly AI generation credits. We hit that ceiling often enough that we are now on Pro at €12 a month, roughly $14. The connector never changed. Our volume did.

Shopify belongs here too, and it is the one people misread most often. The connector is free and official. It is also useless without an active store plan, because an empty store has no orders for Claude to read.

The disguised version is the one that catches careful people. Perplexity’s connector runs on the Sonar API, billed per use and entirely separate from a Perplexity subscription — paying $20 a month for Pro does not make the MCP free. It buys you $5 a month of API credit, and that is the whole overlap between the two products. Neither vendor is hiding anything. But when a connector sits inside something you already pay for, nobody thinks to check whether it is included.

When you evaluate a Pattern 3 connector, the question is never “what does the MCP cost.” It is “what is the cheapest plan of this product that makes the MCP worth having.”

Pattern 4 — Metered

You pay per unit, and the unit is invisible until it isn’t. X is the version most people meet first: pay-per-use API, and since February 2026 no free lane at all for new developers. Automation platforms like Zapier and Make meter tasks and operations, and the meter moves every time a workflow runs, whether or not you were the one who triggered it.

The trap inside this pattern is that metered costs can hide behind a free connector. Windsor.ai looks like a way to read X data without paying for X. It isn’t. Connecting X through Windsor asks you for four X developer credentials, which means you are supplying your own paid X API access and Windsor is riding on top of it. We found that out by trying it. A free connector stacked on a metered API is still a metered API — the bridge does not absorb the toll.

Metered pricing is not bad pricing. It is the only pattern where the bill reflects how much you actually used, which is fair. It is also the only pattern where the bill can grow while you sleep, and that earns a ceiling before it earns your trust.

Five traps

1. The day-31 cliff

You build during a trial, you plan around trial capacity, and the plan quietly stops being possible a month later.

The complication is that day thirty-one is not a fixed date. It is a date the vendor can move. Our Airtable trial expired on schedule and the workspace dropped to the free plan on its own: no warning, no charge, nothing broken, just a small grey badge where the plan name used to be. Later the same day, an email offering to extend the trial by two weeks. We declined, because we were not using the paid features. But the mechanism is worth naming — the cliff you wrote in your calendar is not necessarily the cliff you get, and an extension resets the date in your head without resetting the risk.

Note also what that downgrade cost us, which was nothing. Two small bases, well inside the free ceiling, still working. That is the honest shape of this trap: it hurts in proportion to what you built on top of it. Someone running live automations wakes up to stopped automations.

The Windsor cliff above added a second lesson we didn’t expect: the way down is one-way. Once we were over the limit, deselecting connected accounts to get back under it worked cleanly — nothing was revoked, and the account list stayed populated. Selecting them again did not. The counter refused to climb, and the popup explaining why offered a subscription instead. So the choice on day thirty-one isn’t “pick what to keep” — it’s “come down to one, and whatever you give up stays given up until you pay.”

A footnote on vocabulary, because it cost us a minute of real confusion: on that same screen, the popup called our account a trial, the plan badge called it Free, and the API called it FREE. Three names for one state. Trust the counter, not the label.

What to do: the day you connect anything with a trial, write the expiry date somewhere you will actually see it — not inside the tool, which is the thing that will stop working. If the vendor extends it, move the date instead of deleting it. Then answer one question in advance: if this drops to the free tier tomorrow, what breaks?

2. The free connector on a paid product

Pattern 3 in practice. You read “MCP included at no additional cost,” you connect it, and the real price of the setup turns out to be the subscription you now have to keep.

What to do: price the cheapest plan that makes the connector useful, not the connector.

3. Metered platforms: the meter counts steps, not work

When this guide first went up, this section said that one instruction to Claude can become dozens of operations on an automation platform, and promised a measurement instead of a guess. Here is the measurement — and it corrects us.

On 1 August 2026 we built scenarios in Make on the free plan and read the credit counter after every run. Three things held.

Building costs nothing. Designing a scenario and switching it on moved the counter by zero. You can spend an afternoon assembling something elaborate without spending a credit. The meter starts when the thing runs.

A run costs one credit per module, minus one. A six-module scenario charged five. The final step — the one that hands the result back to Claude — was free. So the number to estimate is not “how often will this run,” it is “how many steps does it touch, multiplied by how often.”

The size of the job did not move the meter. We pointed a scenario at five records expecting the loop to charge for each one, and we wrote the expected number down before running it — sixteen — because a prediction made after the fact is not a prediction. The counter moved by five. The loop step itself charged once, no matter how many items passed through it.

One boundary we have not tested, and it matters: our scenario looped over data it already had. We have not measured what happens when the step after a loop makes a real external call, one API request per record, which is where the multiplication everyone warns about would actually appear. Treat “loops are free” as unproven. Treat “the meter counts steps, not megabytes” as measured. And the Zapier half of this comparison is still missing — Zapier bills per task rather than per module, so the arithmetic is not transferable, and we are not going to publish a number we have not seen.

A naming trap while you are here: Make renamed its billing unit from operations to credits in August 2025. Most guides you will find still say operations, including older ones of ours. The accounting is the same, the vocabulary is not, and it makes older cost estimates hard to compare.

What to do: before pointing Claude at a metered platform, turn on the per-module credit display in the canvas settings — in Make it is off by default, which is a strange default for a product that bills by the module. Then run one small task and read the counter afterwards. You are not testing whether it works. You are calibrating what a sentence costs.

4. Pay-per-use without a ceiling

Usage-based pricing has no natural stopping point and no built-in “that’s enough.”

Read the rate card for the thing you actually intend to do, not the headline number. Publishing to X costs around a cent and a half per post, which sounds like nothing — until you notice that a post containing a link costs about $0.20, roughly thirteen times more, after a repricing in April 2026. If your automation exists to share links, and most creators’ automations do, you are not on the number you think you are on.

Not every metered product behaves this way, and the difference is worth checking before you worry. Make sells no overage below its Enterprise tier: when the month’s credits are gone, scenarios stop until the cycle resets. That is a hard stop, not a bill — irritating on a Tuesday morning, but it cannot surprise you at the end of the month. The API you need to cap is the one that will happily keep serving you past the number you had in mind.

What to do: set a spending cap in the vendor’s dashboard before the first call. Not after the first bill.

5. Auto-recharge — the cost that is on no pricing page

This is the one we learned the expensive way, on our own account, in July 2026.

Credit-based billing systems frequently ship with auto-recharge switched on by default and no cap set. The balance never runs out, so nothing ever fails, so nothing ever tells you that your consumption has changed. The mechanism designed to prevent an interruption is the same mechanism that prevents you from noticing.

What to do, before you connect anything metered: open billing settings, turn auto-recharge off, and set a hard monthly limit. If you would rather leave auto-recharge on, set the cap anyway. It takes ten seconds and it is the single highest-value thing on this page.

Three stacks, three monthly bills

Patterns are only useful if they add up to a number. Here is what three realistic setups actually cost. We pay in euros with EU VAT included, so the dollar figures are approximate conversions of real invoices, not list prices.

The solo creator stack has everything Claude needs to draft, organise, schedule and file, and it costs nothing. The only reason to spend is Canva Pro, and only once you are hitting the AI generation ceiling often enough to be annoyed by it.

The e-commerce operator adds Stripe and Shopify. Both connectors are free, you are already paying for the store, and Stripe takes its cut per transaction. No new subscription enters your life because of MCP.

The marketer is where “my AI setup is free” stops being true. Windsor’s paid tier sets the floor at $19 a month if you commit annually, or $23 month to month; Supermetrics sets the ceiling at $55 a month, or $44 if you commit annually. It is worth knowing in advance that this jump is caused by analytics, not by Claude.

The one rule

Before connecting anything, answer two questions.

Which of the four patterns is this? And what happens on day thirty-one?

If you cannot answer the second one, you have not finished evaluating the tool.

You Might Also Like

More from our guides on connecting Claude to the tools you already pay for.