Can an agent pay for a research service?

Yes, and research is the cleanest case for an agent paying per call, because research is the thing agents do most and the thing that is most obviously wasteful to buy in bulk. A research provider metered by the query costs a cent when the agent needs one answer and a dollar when it needs a hundred, and the agent knows which before it starts. A research subscription costs the same whether the agent ran one query this month or ten thousand, and it required a person to open an account, attach a card, copy a key, and hope. This post is one real research run, what it cost at each step, and the controls that keep a long loop from becoming an expensive one.

The providers

Through one MCP server, an agent reaches Exa, Parallel, Tavily, Perplexity, Linkup, Valyu, and Brave Search for search and deep research, and Firecrawl, Jina, Apify, and the rest of the scraping pillar for reading what the search found. They differ in shape. A semantic search returns ranked results with snippets for around a cent. A page read returns the content of one URL for around a cent. A deep research task, which plans, searches, reads, and writes a sourced answer, is ten cents at the standard tier and thirty cents at the deepest. The agent does not memorize this. It asks.

One run, priced

The task: a sourced summary of how Stripe's Machine Payments Protocol differs from x402, with primary sources.

Step one, the plan. The agent calls consult with that intent. What comes back is a set of options with prices, and it looks roughly like this: a chain of semantic searches at a cent each followed by page reads at a cent each, which for this question will be six to ten cents and give the agent the raw material; or a single deep research task at ten cents that does the whole thing and returns a synthesized answer with citations; or the deepest tier at thirty cents for a question that needs many hops. The agent chooses, and the choice is a cost decision made with the prices in front of it, which is a decision no agent with a subscription ever gets to make.

Step two, the search. Suppose it takes the chain. The first call:

{
  "service": "exa",
  "action": "search",
  "params": { "query": "Stripe Machine Payments Protocol Tempo x402 differences", "numResults": 10 },
  "max_cost_cents": 2,
  "intent": "primary sources on MPP vs x402"
}

The price is at or under two cents, so it runs, and the agent gets ten ranked results and a receipt showing what was charged, on which key, and the transaction it traces to. Three more searches with narrower queries, four cents so far.

Step three, the reads. The agent picks the four most useful URLs, the Stripe MPP documentation, the x402 introduction, and two others, and reads each one at a cent. Eight cents so far, and the agent has the primary sources in hand.

Step four, the deep task, if needed. If the chain had not been enough, the agent would have escalated:

{
  "service": "parallel",
  "action": "task",
  "params": { "input": "Compare Stripe MPP and x402: settlement, buyer requirements, fees, status. Cite primary sources.", "processor": "pro" },
  "max_cost_cents": 10,
  "intent": "sourced comparison, one shot"
}

Ten cents, one call, a written answer with citations. The agent did not need it this time; the chain was enough. Total for the run: eight cents, every step priced before it ran and billed only on success.

That is the shape of research paid per call. The agent spent what the question needed and nothing more, it never opened an account with any of the four providers it touched, and there is no key for any of them anywhere in its environment.

Keeping a long loop inside a budget

Research loops are the case where an agent's spending runs away, because the natural structure of research is "search, read, realize you need another search". Three controls bound it, and they stack.

The agent's key carries a monthly limit. A research agent with a twenty-five dollar month can run roughly two thousand cent-level calls, or two hundred deep tasks, before the next paid call is refused with the reason and nothing charged.

The agent states a maximum cost on every call. This is what stops the single expensive step: an agent that meant to run a ten-cent task and asked for the thirty-cent tier by mistake is refused if its ceiling was ten, and gets the actual price back so it can decide.

The key is restricted to the categories research needs. A key limited to Search and Data cannot start a sandbox or generate an image, no matter what a page it read suggested. For a research agent that reads the open web all day, this is the control that matters most, and it costs nothing to set.

Every call, every price, and every refusal is in the agent history, attributed to the key. When a loop does something surprising, the history is where you find out what it tried.

Why per call is cheaper for this

An agent's research demand is bursty. It runs nothing for a day and then two hundred queries in an hour when a task lands. A subscription prices the peak and charges it every month. Per-call pricing charges the burst when it happens and nothing when it does not, and it lets the agent pick the cheapest adequate provider for each step rather than the one it has a key for. The economics only work because the calls are settled on machine rails, x402 in USDC on Base, Stripe's Machine Payments Protocol, or Tempo, where a one-cent payment carries almost no overhead. On a card network the same run would cost more in fees than in research.

The search and scrape workflow is written out end to end in the search and scrape recipe. The controls above are set the way described in how to give an AI agent a budget.

npx @vaaya/mcp install

Questions

Which research APIs can an AI agent pay for?

Through Vaaya: Exa, Parallel, Tavily, Perplexity, Linkup, Valyu, and Brave Search for search and deep research, plus Firecrawl and others for reading the pages found. Each is metered per query or per task and priced before the call runs.

How much does a research call cost?

It depends on depth. A semantic search is around a cent. A page read is around a cent. A deep research task that plans, searches, reads, and synthesises is ten cents at the standard tier and thirty at the deepest. The agent sees the exact price in the plan before it commits.

Can I cap what a research agent spends?

Yes, three ways at once. A monthly limit on the agent's key, a maximum cost the agent states on every call, and a category restriction so the key can only spend on Search and Data. A call above any of those is refused before it runs and costs nothing.