nvNeuralVerge

One API for AI Agents: The Data Layer vs. Stitching Five Vendors

One API for AI agents: research, extraction and enrichment on one account and one response shape, reached over REST or as MCP tools on a single URL.

By the NeuralVerge team · 15 min read · Updated October 2026

▸On this page

A one API for AI agents is a single account that covers the three things an agent keeps needing: research for questions it cannot answer yet, extraction for pages and documents it can point at, and enrichment for facts about a specific company or person. You reach it over REST from code or as tools on one hosted MCP server. This article is about what one account and one response shape change, not about the cost model, which has its own article.

The short version

  • —The answer. If an agent workflow will need a second kind of data, one account with one response shape saves you a vendor signup, a key, a parser and a failure mode per capability. If it will only ever need one, a single-purpose tool is a fine choice.
  • —What you get. AI research, AI extraction, Search and a catalog of data sources under one key. Search, extraction and every source are also MCP tools. AI research runs over REST only.
  • —Who it is for. Teams building sales, recruiting, diligence or research agents that cross domains, and lean teams that would rather maintain one integration than five.
  • —What it costs. Points from a monthly plan, from $50 for 50,000 points. Search and AI Extract are 5 points a call, research is 20 to 400 points by depth tier, and each source has its own price. The pricing page is the source of truth.
  • —The tradeoff. You give up picking the single best vendor for each capability in exchange for fewer moving parts. How much that is worth depends on your workflow, and the last sections give you a way to check.

What the phrase means

It is easiest to place against three things teams do instead.

Versus stitching point solutions together

The common path is a specialist per job: a search-grounded answer API for research, a scraping or extraction API for turning pages into fields, an enrichment API for company and person data. Each is often genuinely good at its one thing, which is usually why it was picked. What it adds is everything around the capability: a separate signup, a key to store and rotate, a rate limit to track and a response shape to write a parser for. None of that shows in a demo of one vendor. All of it shows up once a pipeline depends on three of them working at once.

The money side of this, including a cost model you can fill in, is in the cost of stitching five APIs together. This article takes the other half: what the agent and the code look like when the seams are gone.

Versus building it in-house

The other path is skipping vendors for one capability and building it, usually extraction, because scraping looks approachable. That holds until the target site changes its markup, a page needs rendering, or a new source needs a new parser. Scrapers and parsers are a standing engineering commitment for something that is not the product you are building, and the cost is easy to underestimate the first time and hard to ignore by the third layout change.

Versus a catalog on its own

A data catalog is one request style across many lookups: registers, profiles, marketplaces, contact data. That is the enrichment third of this picture, and the data catalog article covers how it works, what a real response looks like and what it costs. A catalog does not answer an open question that has no obvious source, and it does not read a page it has no lookup for. One API for AI agents is the catalog plus the two capabilities that cover those gaps, on the same account.

How the three capabilities fit together

It rarely looks like a three-capability problem at the start. A team picks up an extraction API because the first task is "turn this page into fields". Then the workflow needs a fact it has no URL for yet, which is a research problem. Then it needs a person's current title or a work email, which is enrichment. Each capability answers a different kind of question.

Research: you have a question, not a source

Research answers "what is the answer to this", when you do not hold a specific page or document. A run plans the lookups, checks structured sources first, searches the open web, cross-checks what it finds and returns an answer with a citation behind each claim. It is asynchronous: you start a run, get a session_id back, and poll for the result. It is priced by depth tier. The request and the status poll are in the documentation.

Extraction: you have the source, you want fields

Extraction answers "get me these exact fields", when you already know the URL or the document. You pass the URL and plain-language instructions, optionally with a JSON Schema, and get structured fields back instead of markup. It is a flat 5 points a call.

Enrichment: you have an entity, you want what is missing

Enrichment answers "fill in what we do not know about this company or person", against dedicated sources in the catalog: professional-network profiles, startup and funding data, company registers, email and phone lookups. Each source is its own entry with its own input, fields and price.

Which capability answers which question

CapabilityYou start withYou getPoints
AI researchA question in plain languageA cited answer20 to 400, by depth tier
SearchA queryRanked results to read or hand on5 per call
AI extractionA URL or document, and instructionsStructured fields5 per call
A data sourceA name, domain, registration number or similarTyped fields for that sourceSet per source, shown on its page

A pipeline that only ever needs one row can reasonably use a single-purpose tool. The moment a second or third row shows up, and for most agent workflows it does, you are choosing between a second vendor integration and an account that already covers the new question.

A worked example: qualifying an inbound lead

Take an illustrative case. An agent has to decide whether a new signup from Acme Oy (Finland) is worth a sales follow-up, using only the company name and a work email from the form. Nothing here is a claim about a real company.

  1. —Enrichment resolves the email's domain to a company and runs a company lookup on it: industry, headquarters and a short description, a quick signal of whether Acme Oy fits the target segment at all.
  2. —Research answers what enrichment cannot: whether Acme Oy has raised funding recently, whether its registered structure looks legitimate, and whether anything public raises a flag. That is a compound question that benefits from planning, several sources and cross-checking, not one lookup.
  3. —Extraction pulls structured fields from Acme Oy's own pricing or product page, if the agent needs to know what the company sells before deciding how to position outreach.

On three separate vendors that is three keys, three response shapes to normalize into one lead record and three places something can fail without anyone noticing. On one account it is three calls against the same key. The agent's own logic, or a person reviewing the result, combines them into one decision without a parser per step. A rough cost for the run: 100 points for research at the Core tier, 5 for the extraction, and the price shown on the page of whichever company source you use.

What changes with one account

One set of credentials

Three vendors means three signups, three keys to store and rotate and three usage limits to watch separately. One account collapses that to a key and a dashboard. It is a small thing on day one and a real reduction in silent breakage later: an expired trial or a missed rotation on one of three vendors can take down a third of a pipeline with nothing pointing at which one.

One response shape for the parts that repeat

This is the part that shows up in code. Three vendors define three JSON shapes, with different names for a citation, different conventions for a missing value and different error formats. Every difference is a parser you write and keep working while each vendor changes on its own schedule.

On NeuralVerge, the calls you make share the same outer structure: a session_id, a kind, a human text rendering and a machine object holding the typed result. Search, extraction and the data sources return that structure directly; a research run returns it nested under results once you poll for it. The fields inside machine still differ by capability and by source, as they should, because a company register and a funding profile do not hold the same facts. What you stop writing is the outer parser, the error handling and the audit trail, once per vendor. The data catalog article shows a captured response in full.

One billing unit

Every call is charged in points from the same monthly allowance, so one dashboard shows what research, extraction and enrichment cost together. That makes it possible to answer "what does one qualified lead cost us" with a single number instead of adding three invoices in three currencies and billing cycles.

One tool surface for agents

Calling a capability from backend code and exposing it as a tool an agent picks up mid-task are the two ways any of this is used. With three vendors an agent that needs research, extraction or enrichment is choosing between three tool definitions, three schemas to reason about and three response shapes to parse. With one platform, Search, AI Extract and every data source are tools on one hosted MCP server, each returning the same shape as its REST endpoint. The agent sees one consistent tool surface.

REST and MCP

Everything above can be reached two ways. Over REST, your code calls an endpoint when it knows a call is needed. Over MCP, an agent is handed the capabilities as tools and decides mid-task which to use. The same thing is reachable over both, with one exception that matters for this article.

CapabilityRESTMCP tool
AI researchYes, run-research then a status pollNo, REST only
SearchYes, run-searchYes, run_search
AI extractionYes, run-extractYes, run_extract
Each data sourceYes, one endpoint per sourceYes, one tool per source, such as run_email_finder

REST paths keep hyphens and MCP tool names use underscores. Research is REST only because a run is asynchronous and an MCP tool has to return inside the call. The usual shape is a backend that starts research over REST while the agent uses Search, extraction and the sources as MCP tools around it, all from the same account.

The same call, both ways

This is the extraction request from the AI extraction product page, with the fictional Acme example used across the site. Over REST:

curl https://api.neuralverge.ai/functions/v1/run-extract \
  -H "Authorization: Bearer <API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://example.com/company/acme",
    "instructions": "Extract the company name, founding year and employee range.",
    "settings": { "country_code": "us" }
  }'

From an agent over MCP, the tool call carries the same arguments:

{
  "tool": "run_extract",
  "arguments": {
    "url": "https://example.com/company/acme",
    "instructions": "Extract the company name, founding year and employee range.",
    "settings": { "country_code": "us" }
  }
}

The product page shows this response for the call. It is an illustration with a fictional company, not a recorded run:

{
  "session_id": "11111111-1111-1111-1111-111111111111",
  "kind": "extract",
  "url": "https://example.com/company/acme",
  "machine": {
    "company_name": "Acme",
    "founded": "2019",
    "employees": "50-100"
  }
}

Both routes return the same response shape and are priced the same, with no separate MCP rate. That is what makes the one-account argument hold for agents: the tool surface is the catalog, not a second integration. Exact parameters and response schemas are in the documentation, so check them before you build.

Connecting a client

NeuralVerge hosts the MCP server at https://mcp.neuralverge.ai/mcp. Add that one URL to Claude, ChatGPT, Cursor, VS Code, Gemini CLI or any other MCP client, sign in through the browser and click Allow. There is no key to paste into a config file and nothing to build or host. Per-client steps are in the connection guides, and the MCP server guide walks through a full connection. Connected clients appear under Settings, API in the dashboard with a "Connector:" label and can be revoked there.

Choosing between them

Use REST for scheduled pipelines and anything where your code already knows when to call, and for research in every case. Use MCP when the agent is picking tools as it goes. Most real setups use both.

Cost and limits

Every call is billed in points from the monthly plan. Plans run from Base at $50 for 50,000 points, through Core at $100 for 100,000, Pro at $250 for 250,000 and Ultima at $500 for 500,000, to Enterprise at $1,000 for 1,000,000, with a custom tier above that. Search and AI Extract are 5 points a call. Research is priced by depth tier:

  • —Lite: 20 points, aimed at 30 to 90 seconds.
  • —Base: 50 points, 1 to 2 minutes.
  • —Core: 100 points, 2 to 4 minutes.
  • —Pro: 200 points, 3 to 7 minutes.
  • —Ultima: 400 points, 5 to 12 minutes.

The times are the estimates shown on the pricing page, not guarantees. Data sources are priced per entry, often per returned record or per lookup, and the figure on each source page is the one to use. In an MCP session the price is the same as the REST call.

A few limits worth knowing before you commit:

  • —Empty calls can still cost points. A source's billing line says whether a miss is free. For extraction, a call on a URL that returned nothing was still charged 5 points in testing, so budget for some calls not returning what you hoped.
  • —Depth varies by capability. A bundle can be thinner than a specialist on one dataset. Check the fields you need against the source page rather than the category name.
  • —Swapping one piece is harder inside a bundle. The capabilities share account infrastructure that a single-purpose vendor does not, so replacing one later is a bigger job than replacing one stitched vendor. The counterweight is the upkeep the stitched version costs on every day it runs.
  • —The catalog changes. Sources are added and fixed over time, so treat any count as a snapshot and check the source catalog for the current list.

How to decide

Run these questions against the workflow you actually have.

  • —Will the workflow need more than one kind of data? Count the capabilities in your real task, not your first demo. One means a specialist is fine. Two or three means bundle.
  • —Does the bundle cover each capability or just the category label? Compare the fields you need against the source page for each one.
  • —Is the response shape the same across them? Send one real request to two capabilities and compare the outer structure directly instead of trusting a features page.
  • —Do REST and MCP return the same thing for the same capability? If not, that is two integrations to keep in sync, not one.
  • —Does an agent choose its own tools? If yes, the single MCP URL is the strongest argument for consolidating. If your code decides everything, REST alone covers it.
  • —Who does the work? Sales and recruiting agents researching a company, enriching a contact and reading a source page in one flow are the clearest fit. So are diligence and KYB pipelines that combine research for compound questions with registry and identity lookups, and lean teams that would rather maintain one integration than a growing list.
  • —Are you optimizing for the best single tool or for fewer parts? Both are defensible. Pick on purpose and write down which.

A faster test than any comparison is to run one workflow that touches more than one capability through a candidate and count how many response shapes it took to reach the final answer.

Frequently asked questions

Doesn't a best-of-breed vendor per capability still perform better than a bundle?

Sometimes, on a single capability in isolation. The tradeoff is what it costs to get there: separate contracts, separate auth, separate response shapes, and separate failure modes to monitor. A bundle is worth it when the combined maintenance cost of running several point solutions outweighs whatever edge any one of them has on its single strength — which is common, but not universal, so it's worth checking against the specific capability that matters most for a given workflow.

What happens if I only need one capability today, like extraction?

Nothing forces you to use research or enrichment just because they're in the same catalog — call only the endpoint the workflow needs today. The advantage shows up later, if the workflow grows to need a second capability, because it's already inside the same account, auth, and response shape rather than a new vendor to onboard.

Does bundling mean giving up flexibility to swap out one piece later?

It's a real tradeoff to weigh, not something to wave away. Swapping one piece of a stitched pipeline for a different vendor is usually easier than swapping out one capability inside a bundle, since a bundle's capabilities share account infrastructure a single-purpose vendor doesn't. The counterweight is how much maintenance the stitched version costs every day it's running, not just how easy it is to change later.

Is REST or MCP the "right" way to call a bundled API?

Neither is inherently right — they're two ways to reach the same platform. REST fits code that already knows when a call should happen, and it covers everything, including AI research. MCP fits an agent that needs to decide for itself, mid-task, which tool to reach for: Search, AI Extract and every data source are MCP tools; AI research runs over REST. Each MCP tool returns the same response shape as its REST endpoint and is priced the same, so switching between them isn't a second integration.

Can I use one API for AI agents over MCP, or only through REST?

Yes, for most of it. NeuralVerge hosts an MCP server at https://mcp.neuralverge.ai/mcp. Add that URL to Claude, ChatGPT, Cursor or any MCP client, sign in through the browser and click Allow, and Search, AI extraction (run_extract) and every data source appear as tools. AI research is the exception: it runs over REST, so a typical agent calls research over REST and uses the MCP tools around it. MCP calls are priced the same as REST and return the same response shape.

About NeuralVerge

Give your agents structured, cited, real-world data from 300+ sources and the open web — through one API or MCP server.

AI Agents & MCP on the NeuralVerge blog.