Web search MCP servers you can pay per call

What web search through MCP APIs do

Web search through Model Context Protocol connects an agent's tool interface to a search operation. MCP describes how a client discovers and invokes tools. The underlying paid product still has an output contract: search results, a grounded answer or a research response. A service mentioning MCP in its listing does not by itself prove that a specific search route is exposed as an MCP tool or that its transport accepts the client's payment workflow.

This page starts with the existing web-search MCP subset and then checks documented search endpoints. It compares the paid search products of listings that advertise MCP, while retaining the distinction between a catalog claim and a tested integration. Before connecting an agent, inspect the provider's MCP URL, transport, tool names and argument schema. Confirm where payment is requested and how the selected client handles that request.

A normal search request and an MCP tool invocation can ultimately reach the same provider operation, yet the integration has extra steps. The client may discover tools, select one, validate arguments and map the tool result into a response to the user. Those steps need explicit errors and spending controls. A successful catalog HTTP check is evidence about service monitoring rather than proof that this complete tool flow works.

A common use is an assistant that needs current sources during a conversation. Another is a research agent that repeats a query with a revised scope. In either case, tool availability should not become an instruction to search indefinitely. Set the allowed products, maximum calls and acceptable evidence before handing the agent a search tool, and keep the user's research question attached to the tool arguments.

Which outputs the documented endpoints provide

The result types below describe what the paid search endpoint advertises, not the protocol used to invoke it. Search results, synthesized answers and specialized discovery belong in separate groups. A tool-definition response describes available actions; it is not the result of a search. Likewise, an endpoint selling a skill document may help configure an agent without returning live web evidence.

A gateway's MCP label can apply to its broader tool collection. Endpoint selection still excludes page fetches, internal record searches, unrelated utilities and instruction files. The listed price comes from a matching search route and data type. It is not inferred from the gateway's cheapest tool or from a transport's existence.

Each price is the cheapest endpoint of that service for that kind of data: the same number wherever this page quotes it.

To assess a real integration, read the provider's current client instructions and test the intended tool with a representative query. Keep the MCP tool name, its arguments and the underlying resource or product together in the execution log. If the provider returns a job identifier, distinguish submitting the job from retrieving its final research result. A protocol connection and a completed search answer are separate milestones.

Compare 3 web search through MCP services

One row per service, ordered by observed paying wallets. The quote belongs to the endpoint named in that row. Primary data types determine the row quote when available; adjacent products keep their own type prices below.

$/CALL is the price of the endpoint named under it: the cheapest ranked web discovery results endpoint of the service or, for a service without one, its cheapest web search endpoint. Never the cheapest tool of a gateway. Other kinds of data have their own price in the tables below. WEB SEARCH ENDPOINTS counts the web search endpoints out of all the endpoints the service lists. BUYERS 30D counts distinct wallets that paid the whole service on-chain in 30 days, not only its web search endpoints. "Generalist" marks a gateway that sells more than three unrelated kinds of tools: it is never called the cheapest.

How to choose a web search through MCP endpoint

Check the tool boundary

Inspect what the server advertises through tool discovery and compare the argument schema with the documented search route. Determine whether the operation returns a result immediately, streams progress or starts a job that needs another call. A transport connection without the intended search tool does not satisfy the task. Pin the server configuration your client actually uses so a later tool change can be diagnosed.

Place payment in the client flow

Confirm how the chosen client surfaces HTTP 402 terms and who controls the wallet or signing step. Keep the tool arguments bound to the requested resource and amount; a server description does not establish that a particular MCP client can perform that payment automatically. Test the unpaid challenge and error path before making the tool available to an unattended agent. Use a bounded retry policy for network failures and ambiguous job completion.

Evaluate the evidence after invocation

A tool result still needs source review. Preserve URLs, query settings and observation time, and distinguish provider-generated summaries from material retrieved from a source. Limit the number of tool calls allowed for one answer and count subsequent page reads separately. A well-formed MCP response can contain inadequate evidence for the user's question; decide that from the content, rather than from protocol success.

Match the request to the output

Start with one representative input and a written acceptance criterion. A successful HTTP status is not enough: inspect the fields your application needs, the distinction between missing and negative results, and the provenance supplied with the response. Keep a small reviewed evaluation set that includes ambiguous inputs as well as straightforward examples. Changing a request parameter can change the product and its payment requirement, so preserve those parameters alongside the endpoint name.

Check the operating evidence

The table reports service monitoring, with a median uptime of 100% and a median selected endpoint price of $0.01. Monitoring latency describes our checks, rather than a benchmark of a complete paid workflow. A request that triggers a larger job may have a different duration. Compare a provider's documented limits with the concurrency, timeout and result size your caller can handle. Record errors separately from empty answers so an outage does not look like a meaningful absence.

Plan changes and retries

Cache according to the provider's terms and the freshness your task needs. Keep a versioned record of input normalization and response mapping so a later provider change can be investigated. Before an automatic retry, decide whether the previous request could have completed and whether repeating it incurs another charge. An agent should have a request budget and an explicit stopping condition, especially when one answer causes follow-up calls. Validate required fields before submitting a payment rather than discovering an invalid input through a long retry loop.

Choose by the job your caller needs to complete

Start with the result your next step can consume. The candidates below come from the same data-type classification as the comparison table. The displayed job quote is the lowest matching type price for that job, with its source endpoint retained.

Ranked web discovery results

Compare the documented input, returned fields and paid unit for this output. Preserve ambiguity and missing-result states in the caller.

Candidates today, most buyers first: vaaya ($0.01/call, generalist, 57 buyers) · Tanod ($0.007/call, generalist, 26 buyers) · agdata ($0.02285/call, generalist, 19 buyers)

Search with returned page text or passages

Compare the documented input, returned fields and paid unit for this output. Preserve ambiguity and missing-result states in the caller.

Candidates today, most buyers first: agdata ($0.022545/call, generalist, 19 buyers)

News discovery or explicit news mode

Compare the documented input, returned fields and paid unit for this output. Preserve ambiguity and missing-result states in the caller.

Candidates today, most buyers first: agdata ($0.0239/call, generalist, 19 buyers)

What the providers describe

The descriptions below attribute advertised behavior and limitations to the provider. They were reviewed against its public listing and endpoint descriptions captured on October 21, 2026; operating figures and prices are generated from the page model.

agdata

AGData separates Google search results from page-reading search. Its Google discovery route costs $0.02285, while /x402/web-search costs $0.022545 and advertises search with page text. The description bounds that reading operation to a small set of pages and a per-page text limit; it is not an unrestricted crawl. A separate Google News route costs $0.0239. Choose the output before comparing the quotes: a snippet list, retrieved page text and news discovery carry different downstream evidence and request limits.

3 web search endpoints of 22 · $0.02285/call (GET /x402/google-search) · Base/POL/SOL · uptime 30d 100% · 57 ms · 19 buyers in 30 days

Tanod

Tanod describes /v1/search as independent web-index search with ranked titles, URLs and snippets at $0.007. Its advertised query controls include result count, country and freshness. That supports discovering pages for a follow-up read, while a snippet alone may omit the evidence needed for a final answer. The listing also advertises MCP access; verify the intended tool schema and payment flow in the client before treating the catalog's HTTP route as a completed MCP integration. Save query settings and returned URLs together.

1 web search endpoint of 6 · $0.007/call (GET /v1/search) · ARB/Base/POL/SOL · uptime 30d 100% · 464 ms · 26 buyers in 30 days

vaaya

Vaaya describes an MCP server that brokers paid API operations, including an Exa semantic-search route at $0.01. The documented HTTP operation accepts a result-count setting and describes a people category for lead discovery. This is an upstream product claim, rather than evidence that a particular client can invoke the tool and authorize payment. Inspect the live MCP tool name and argument schema, bind the request to the intended product and keep the returned URLs for source review. Its separate memory-search operation queries stored user memories; that result is not live web discovery and does not supply this page's quote.

1 web search endpoint of 7 · $0.01/call (POST /api/run/exa/search) · Base · uptime 30d 100% · 450 ms · 57 buyers in 30 days

Cost per call and at larger volumes

The median selected endpoint costs $0.01 per call. At an unchanged request shape, 100,000 calls total $1,000.00. The table multiplies reference quotes so a prototype and a larger workload can be compared on the same basis.

Each service at the price of its main web search endpoint, the one named in the table above.

The paid unit matters. One request may return a bounded list, inspect one input, start a job or retrieve an existing result. Check that unit before multiplying a quote by your business volume. A second product, a deeper lookup or a retry can add a charge, while no-match and failure handling depend on the provider's terms. A catalog reference quote does not establish the amount for a larger batch or a changed request body.

For a production budget, measure the calls required for an accepted result. Save the endpoint, normalized request, quoted amount and outcome in the job record. This lets you distinguish a low submission price from a low effective cost. Read the payment requirement for the exact input when the provider describes variable pricing; do not substitute a gateway's cheapest unrelated tool.

What the captured 402 establishes

The catalog lists Tanod at GET /v1/search, but the stored 402 advertises POST with a JSON body containing query and count in extensions.bazaar.info.input. The catalog method and captured input schema disagree. The payload's error message also says the route takes POST with JSON and that the POST challenge quotes the exact request price. This discovery response therefore cannot substantiate a runnable GET example or the payment terms of a specific POST. Inspect the full response under Payment evidence on the service page; the table's catalog quote is a reference rather than a verified quote for the intended POST body. An MCP tool schema and payment flow still need separate verification. The catalog reference quote for this method and path is $0.007; it is not a verified quote for a different method or request body.

Read the provider's current input schema before building a request. Keep the requested resource, accepted network, asset and amount together when evaluating its challenge. Stored payment evidence does not show a tested paid answer, establish data accuracy or prove that a background job completed. Inspect the full payload on the linked service page and obtain the challenge for the exact request before authorizing payment.

How to interpret this comparison

The listing requires an online, payment-ready service, a usable description, a known price and at least 90% measured uptime over 30 days. Endpoint selection then keeps routes that describe the page's task. A candidate can match a word in its general description yet be excluded because its documented routes return a different product. These are discovery criteria rather than certification of delivered output.

Provider paragraphs attribute statements read in the public listing and endpoint descriptions captured on October 21, 2026. Source, scope and output claims belong to the provider. x402-list monitoring supplies the price and operating figures; it does not independently establish the correctness of the underlying dataset or a vendor's performance claims.

Buyer counts are distinct paying wallets observed for the service in the measurement window, with the coverage limits described on its service page. They do not count people or prove usage of a particular route. Endpoint prices, data-type counts and job candidates share one endpoint classification. If a service disappears, its generated candidates and token-dependent descriptions disappear with it. Open the methodology and linked service pages before deciding what that evidence means for a recurring workflow.

Questions

How to pay

Each service answers an unpaid call with HTTP 402 and its price; your agent signs a USDC payment on the listed network and retries. No API key, no account, no subscription. See what x402 is.