Which MCP servers are people paying for now?

Are people actually paying for MCP servers now?

Yes. People are paying for MCP-powered services today, especially when the server gives an AI agent web data, cloud browsers, authenticated app actions or production infrastructure.

The tricky part is that most companies do not report “MCP revenue” separately. A customer may connect Claude, Cursor or another agent to an MCP server, yet the resulting bill appears as search credits, browser hours, Zapier tasks or tool calls.

We can still see real payment happening. Bright Data currently gives users 5,000 MCP requests free and then charges $1.50 per 1,000 requests. Firecrawl sells 5,000 monthly credits for $16, 100,000 for $83 and 500,000 for $333 on annual billing. Exa includes MCP access with its search platform and charges $7 per 1,000 standard searches, $12 to $15 per 1,000 deeper searches and $1 per 1,000 pages retrieved.

The same pattern appears outside web search. Browserbase charges for the browser infrastructure behind its hosted MCP server. Arcade charges $25 per month on its Team plan, then $0.10 per authorization event and $0.01 per tool call. Zapier counts every successful MCP tool call as two tasks from the customer's existing Zapier allowance.

So paid MCP is already real. What remains much less proven is the idea that thousands of independent developers are earning meaningful subscription revenue from simple standalone MCP connectors.

What are people really paying for when they use a paid MCP server?

People are currently paying for the expensive work behind the MCP call: data, compute, browsers, authentication and reliable execution.

That distinction explains why some MCP servers can charge immediately while others struggle to justify even a small monthly fee.

Take Exa. Its MCP server lets an agent search the live web, fetch pages and run deeper research. A standard search costs $7 per 1,000 requests. Firecrawl sells scraping, crawling and search credits. Bright Data sells web retrieval backed by its data collection and proxy infrastructure. Browserbase puts an actual cloud browser behind the MCP connection.

Compare those with GitHub's official MCP server. GitHub lets an agent read repositories, manage issues, work with pull requests and trigger other GitHub actions, yet GitHub does not sell a separate MCP subscription. The MCP server mainly makes an existing GitHub account easier for an agent to use.

The commercial line is quite clear now. When an MCP call consumes something scarce or expensive, charging for it feels natural. When the server simply translates an agent request into an API call that the customer already paid for, a separate MCP fee is much harder to defend.

MCP product	What the customer is paying for	Current charging model
Exa	Live web search and content	Per request/page
Firecrawl	Search, scraping and crawling	Credit plans
Bright Data	Web data, proxies and browser infrastructure	Per request + infrastructure
Browserbase	Managed cloud browsers	Subscription + usage
Zapier MCP	Actions across 9,000+ integrations	Existing Zapier tasks
Arcade	Authenticated tool execution	Platform fee + usage

Why has paid MCP become much more relevant lately?

Paid MCP matters much more now because the protocol has moved well beyond a Claude Desktop feature and is becoming a common way for agents to reach external tools.

Anthropic said that more than 10,000 active public MCP servers already existed when it transferred MCP to the Agentic AI Foundation. The protocol had by then been adopted across products including ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code.

That changes the commercial opportunity. A server that only worked inside one desktop assistant had a fairly narrow distribution channel. A remote MCP endpoint can now be exposed to several major AI environments without the provider building a separate integration for every one.

The shift toward remote servers matters too. Browserbase, for example, now recommends its hosted Streamable HTTP MCP endpoint rather than requiring users to run the server locally. GitHub also provides a remotely hosted version of its official server. Exa runs a hosted endpoint at mcp.exa.ai.

Once servers become remote services, recurring infrastructure costs and metered usage appear naturally. That is one reason today's most convincing MCP businesses increasingly look like API businesses.

Are web search and scraping the strongest paid MCP category right now?

Yes. Web search, scraping and fresh-data access are currently the clearest examples of people paying through MCP.

Firecrawl is one of the easiest examples to see. Its MCP project has passed 7,000 GitHub stars and 100,000 package downloads, while the commercial service charges for the actual scraping and search workload. Its current plans range from a 1,000-credit free tier to $599 per month for one million credits, before enterprise pricing.

Exa follows almost exactly the same commercial logic. Its official MCP project has around 5,000 GitHub stars, but the more important number is the price attached to the work behind each request: $7 per 1,000 normal searches, $12 to $15 per 1,000 deeper searches and $1 per 1,000 pages for content retrieval.

Bright Data is even more explicit. Its MCP pricing page offers 5,000 free requests, then charges $1.50 per 1,000 requests. A $499 monthly scale plan includes hundreds of thousands of results, while managed stealth-browser traffic is priced separately.

These companies arrived at similar pricing without needing to coordinate with each other. Search and scraping work well as paid MCP products because every additional agent action consumes a measurable resource, and customers already understand that reliable fresh web data costs money.

Are people paying for browser MCP servers too?

Yes. Browser MCP servers have become one of the strongest paid categories because running a reliable browser in the cloud is much harder than exposing a normal API.

Browserbase now runs its hosted MCP endpoint on its own infrastructure and recommends that hosted version for most users. An agent can open pages, click, type, extract information, maintain sessions and use managed browser infrastructure without running Chrome on the user's machine.

The commercial usage behind that endpoint is already substantial. Browserbase says its infrastructure currently handles more than 35 million browser sessions per month. Its customer material names companies including Ramp, Lovable, Clay, Amplitude, Vercel and Microsoft.

The economics are easy to understand. Browserbase's Developer plan includes 100 browser hours, while its Startup tier includes 500. Additional browser time is billed by the hour. Search and page fetching have their own usage prices as well.

Microsoft's Playwright MCP shows the other side of the market. Playwright MCP is extremely popular and can run a browser locally for free. That makes a generic browser wrapper difficult to monetize. Companies such as Browserbase can charge because they handle the cloud browser, concurrency, proxies, identity, recordings and reliability that a local installation leaves to the user.

Are Zapier, Composio and Pipedream becoming the paid “super MCP servers”?

Yes. Integration platforms are turning MCP into one doorway through which an agent can reach thousands of apps.

Zapier now exposes more than 9,000 integrations through MCP. Every successful MCP tool call consumes two Zapier tasks, and those tasks come from the same allowance customers already buy for normal Zapier automation. A sequence that looks up a HubSpot contact and then sends a Gmail message therefore consumes four tasks.

Composio takes a more agent-specific approach. Its current platform includes more than 1,500 toolkits, managed authentication and MCP support. Customers receive 100,000 tool calls per month on the free tier. Pro costs $29 per month, with additional standard calls currently priced at $0.0003 each.

Pipedream Connect follows the same direction. MCP tool calls consume its compute credits, while companies using Connect also pay according to the number of external users connecting accounts.

The important development here is consolidation. A developer building an agent no longer needs to maintain separate Gmail, Slack, Notion, Salesforce and GitHub MCP infrastructure. One integration platform can expose all of them.

That makes these broad platforms much easier to monetize than hundreds of small single-app connectors.

Platform	Breadth today	How MCP usage creates revenue
Zapier	9,000+ integrations	2 tasks per successful tool call
Composio	1,500+ toolkits	Subscription + metered calls
Pipedream	Thousands of apps/actions	Compute credits + external users
Arcade	Large managed tool catalog	Auth events + tool calls

Are companies really willing to pay for MCP authentication?

Yes. Authentication is becoming a serious paid MCP layer because connecting one person's account is easy, while safely connecting thousands of users is a different problem.

Imagine an AI product whose customers want to connect Gmail, Salesforce, Microsoft, Slack and GitHub. Every customer needs the correct OAuth flow, permissions, token refresh, account isolation and revocation. The agent also has to know whose credentials it should use before taking an action.

Arcade charges directly around that problem. Its current Team plan costs $25 per month, with authorization events priced at $0.10 and tool calls at $0.01. Its MCP gateways can authenticate individual users and execute actions without exposing raw credentials to the model.

Composio now sells a similar layer across more than 1,500 toolkits. It handles connected accounts and OAuth while charging for usage above its free allowances. It has also added controls for shared company connections, where several agents or users can act through the same organizational account.

Authentication sounds like plumbing, and it is. In production, though, bad authentication can stop the entire agent product from working or create a serious security problem. Companies are willing to pay to avoid building all of that themselves.

Are businesses paying for MCP security and governance as well?

Yes, although this market is mainly showing up in enterprise deployments rather than individual developer subscriptions.

The security problem grows quickly once a company has dozens of agents and many MCP servers. Someone has to decide which servers employees can connect, which tools each agent can use, what credentials are available and what gets recorded.

GitHub already lets enterprise administrators apply MCP policies and server controls around Copilot. Arcade's enterprise offer includes SSO, role-based access control, audit logs, private registries and deployment inside a customer's own VPC or even an air-gapped environment. Composio offers enterprise controls such as SSO, SCIM, key-management options, IP allowlists and longer log retention.

These products show where MCP spending is going. Early developers cared about connecting a tool. Companies putting agents into real workflows increasingly care about controlling the connection.

The enterprise money should therefore spread beyond individual MCP servers into gateways, policy, credential management and observability.

Why are popular MCP servers from GitHub and other big SaaS companies often free?

Popular first-party MCP servers are often free because the company makes money when the agent uses the underlying product.

GitHub's official MCP server is a good example. Users can connect an AI client directly to repositories, issues, pull requests, Actions and security tools. GitHub provides the server remotely and as open-source software, while the commercial relationship still revolves around GitHub itself and whatever paid GitHub features the customer uses.

The same logic applies across established SaaS. If a customer already pays for a service, charging another $10 simply because Claude or Cursor accesses the account through MCP would add friction to something the provider wants customers to use more.

For a large SaaS company, MCP can therefore work as distribution. More agent activity can make the core account more useful, increase usage of paid features and make the product harder to replace.

This is why counting every official MCP server as a new “MCP business” badly exaggerates the size of the standalone market.

Can an independent developer still make money from a standalone MCP server?

Yes, but the server needs to deliver something harder to copy than an API wrapper.

The best opportunity is a tool with a clear paid outcome. That could be one business lead enriched, one property record assembled, one ecommerce page monitored, one legal database searched or one difficult workflow completed.

Apify gives us a useful glimpse of how this can work. Developers can publish Actors that agents call through MCP and charge per event rather than forcing customers into a monthly subscription. Under Apify's current marketplace economics, the developer receives 80% of qualifying pay-per-event revenue before the underlying platform costs are deducted.

That pricing structure fits agent behavior unusually well. An agent may discover a specialist tool, call it seven times and never need it again. Forcing the customer into a $29 monthly subscription would probably kill the transaction. Charging a small amount for each useful result makes much more sense.

Machine payments are starting to push the idea further. Browserbase already exposes pay-per-use browser sessions through agent-payment protocols, with browser time listed at $0.12 per hour on that gateway. An autonomous agent can therefore buy infrastructure without a traditional account setup.

This part of the market is still early, but it points toward a credible model for independent MCP developers: sell a valuable machine-callable result and let MCP handle discovery and execution.

Does a popular MCP server on GitHub usually mean people are paying for it?

No. GitHub popularity tells us that developers care about an MCP server, but it tells us very little about how much money the server makes.

Microsoft's Playwright MCP makes the problem obvious. It has attracted tens of thousands of GitHub stars and is one of the most visible MCP projects, yet the project itself is free and open source.

Firecrawl's official MCP server has about 7,000 stars, while Exa has around 5,000. Both connect directly to paid infrastructure, so GitHub activity and commercial usage happen to point in the same direction there.

But a company can also have the reverse profile. An enterprise authentication gateway may have almost no public developer community because customers use a hosted endpoint under a corporate contract.

We therefore use repository activity as evidence that developers want the capability, rather than evidence that customers pay for it. Pricing, metered consumption, production customers and underlying infrastructure give us a much stronger read on revenue.

Is simple MCP hosting worth paying for now?

Sometimes, but plain MCP hosting is becoming a commodity surprisingly quickly.

Running a lightweight remote server is cheap. Cloud platforms can host basic HTTP services for very little, and many MCP servers still run locally at no direct cost.

The paid version becomes more interesting when hosting removes operational problems. Browserbase hosts the MCP endpoint and also supplies the browser fleet. Arcade combines hosted gateways with authentication and policies. Composio combines hosting with hundreds of integrations and credential management.

The same logic has shaped MCP marketplaces. Merely listing thousands of servers creates discovery, but the more valuable services sit around execution, credentials, billing and reliability.

We would be cautious about a business whose whole pitch is “we host your MCP server.” Hosting becomes much easier to charge for once something difficult is bundled with it.

Is usage-based pricing becoming the default for paid MCP?

Yes. Usage-based pricing currently fits paid MCP better than the classic per-seat SaaS model.

The pattern appears almost everywhere we look. Bright Data bills MCP requests. Exa charges searches and retrieved pages. Browserbase bills browser capacity, search and fetching. Composio charges additional tool calls. Arcade charges both authorization events and calls. Zapier converts every successful MCP action into tasks. Pipedream meters MCP execution through compute credits.

Firecrawl still sells monthly credit bundles, but even those plans ultimately meter the number of searches or pages an agent consumes.

There is a simple reason. Agent usage can vary enormously without the number of human users changing. A five-person team could run an agent millions of times, while a company with hundreds of employees might barely use one integration.

The bill therefore follows machine work increasingly closely.

Meter	Current examples	Why it fits MCP
Request / search	Exa, Bright Data	Each query creates measurable work
Page / credit	Firecrawl, Exa	Cost rises with retrieval volume
Tool call	Composio, Arcade	Tracks agent actions directly
Automation task	Zapier	Reuses an existing usage meter
Browser time	Browserbase	Reflects compute-heavy execution
External user	Pipedream Connect	Captures multi-user SaaS deployments

Which MCP servers will people probably refuse to pay for?

Generic connectors around easy, public or already-paid APIs are becoming very difficult to monetize.

The supply is simply too large. Anthropic already counted more than 10,000 active public MCP servers before the ecosystem expanded further. Creating a basic server has also become much easier as SDKs, registries and templates have matured.

If an MCP server mainly wraps a public REST endpoint, local filesystem commands or another service's standard API, an open-source replacement can appear quickly. The protocol layer itself offers little protection.

The stronger businesses own something underneath the server. Exa has a search index and retrieval infrastructure. Bright Data has a large web-data and proxy network. Browserbase operates cloud browsers. Zapier has thousands of app integrations. Arcade and Composio handle authentication and execution across many services.

A useful rule emerges from what people are paying for today: the harder it would be for a developer to recreate the underlying capability in a weekend, the easier the MCP server is to monetize.

Which MCP servers are people paying for now?

People are paying for MCP servers now, but the money is concentrated in a few clear categories: web data, cloud browsers, authenticated integrations and production agent infrastructure.

Web search and scraping are the clearest winners today. Firecrawl, Exa and Bright Data all combine active MCP products with straightforward paid usage. Their customers can see exactly what they are buying: searches, scraped pages and fresh external data.

Browser automation is another proven category. Browserbase now serves more than 35 million browser sessions per month across its broader platform and gives MCP users access to the same managed browser infrastructure. The free alternative, such as running Playwright locally, helps clarify why customers pay Browserbase: production users need concurrency, proxies, session persistence, monitoring and reliable remote execution.

Zapier, Composio, Pipedream and Arcade show where a second large pool of spending is developing. Companies want agents that can act across many existing applications without rebuilding OAuth, permissions and execution for every integration.

Generic standalone connectors look much weaker. GitHub itself can give away a sophisticated MCP server because MCP increases the value of its existing product. Open-source alternatives make simple wrappers even harder to charge for.

The current paid MCP market is narrower than the enormous server directories make it look, but much more real than it was a year ago. The businesses getting paid have one thing in common: every MCP call reaches something genuinely valuable or costly behind the endpoint.

Paid MCP category	Examples with clear commercial usage	Our read today
Web search and scraping	Exa, Firecrawl, Bright Data	Strongest standalone category
Cloud browser execution	Browserbase	Clearly monetizable
Broad SaaS actions	Zapier, Composio, Pipedream	Strong and getting broader
Authentication and governance	Arcade, Composio	Stronger as agents reach production
Generic single-API connectors	Many open-source servers	Difficult to charge for
Basic MCP hosting	Numerous providers	Useful, but weak on its own