Connect 359 Sources to Claude in Two Minutes With the NeuralVerge MCP Server

An MCP server for Claude is a remote service Claude connects to so it can call tools mid-conversation. NeuralVerge hosts one at a single URL: add it as a custom connector, sign in through the browser, and Claude can use Search, AI extraction and the sources in the catalog (359 of them at the time of writing, counted from the source pages on this site). No integration code and no key to paste.

The short version

• —What you get: one URL, one sign-in, and Claude can call Search, AI extraction and every data source as tools. AI research is the exception: it runs over REST.
• —Who it is for: anyone who wants Claude to look things up (companies, registries, emails, profiles) instead of working from pasted text. It works on Claude on the web, Claude Desktop and Claude Code.
• —What it costs: the same as the equivalent REST call. The example in this tutorial, email validation, is 1 point, listed on its source page as $1 per 1,000.
• —How long it takes: the title says two minutes. That describes how few steps there are (a URL, a sign-in, a toggle); we have not timed it across setups, so it is not a benchmark.

This is a tutorial, so it is mostly steps. If you want the protocol explained first, read what Model Context Protocol is; for the same pattern applied to one source, see the LinkedIn walkthrough.

Prerequisites

You need three things: a NeuralVerge account, a Claude account (or Claude Code installed), and permission to add connectors. On Team and Enterprise plans an Owner adds the connector first, covered in the next section.

Sign-in is required. A connector without an account will list tools but cannot call them.

What a connector gives Claude

Claude on its own answers from its training plus whatever you paste in. It cannot look up a registry entry, check an email address or read a product page unless something hands it that ability. A connector is that something. Once NeuralVerge is connected, Claude sees a list of named tools, each with a description and parameters, and decides for itself when a question needs one.

The product behind the connector is the platform: research, extraction, and a catalog of data sources covering company registries, company intelligence, startup and funding data, e-commerce, enrichment and social profiles. REST and MCP are two ways of reaching it. For an assistant that must choose a lookup mid-conversation, MCP is the natural one.

What the sign-in does

The server URL is https://mcp.neuralverge.ai/mcp. It uses the Streamable HTTP transport and is the same for every client. Nothing secret is in the configuration.

The client finds out how to authenticate from the server itself, registers, and opens a NeuralVerge sign-in page in the browser. NeuralVerge's documentation calls this OAuth 2.1 with PKCE. The practical result: you sign in as yourself and click Allow. Each connection bills to your organization and appears in the dashboard, next to everything else you run, which also gives you one place to revoke it.

Listing tools works without credentials, so a client can show you the catalog before you sign in. Calling a tool needs the sign-in; without one the server answers 401 and points the client to sign-in.

Connect the MCP server for Claude on web and Desktop

These steps follow NeuralVerge's own setup guide. The wording of Claude's menus is Anthropic's and can change; if a label differs on your screen, the current instructions are at docs.neuralverge.ai.

1. Add the connector

Open Settings → Connectors and click Add custom connector. Name it NeuralVerge and paste the server URL:

https://mcp.neuralverge.ai/mcp

Leave the OAuth fields under Advanced settings empty. Claude discovers the sign-in itself.

2. Sign in

Click Connect. Sign in to NeuralVerge in the window that opens and click Allow.

3. Enable it in a chat

In a new chat, enable NeuralVerge from the + menu under Connectors, then ask a question that needs data. The setup guide's own example is asking for the registered address of Škoda Auto in the Czech company registry.

Other surfaces and plans

The guide says a connector added this way on your Claude account also shows up in Claude Desktop and the Claude mobile apps on the same account. On Team and Enterprise plans, an Owner adds the connector once under Organization settings → Connectors, and members then connect it from their own Settings → Connectors.

Connect Claude Code from the terminal

Claude Code takes the server from the command line:

claude mcp add --transport http neuralverge https://mcp.neuralverge.ai/mcp

Add --scope user to make the server available in every project, or --scope project to share it with your team through .mcp.json. Then start claude, run /mcp, select neuralverge and choose Authenticate. Sign in to NeuralVerge in the browser and click Allow.

Other clients

The same URL works in ChatGPT's developer mode, Cursor, VS Code, Gemini CLI and the MCP Inspector, each with its own config location; the setup guide linked above carries the exact steps for each. Any client that supports the Streamable HTTP transport and MCP authorization can use the same URL.

Check the connection

Verify before you trust it, in two steps.

Ask what it can do. In a chat with the connector enabled, ask Claude which NeuralVerge tools it can see. You should get names such as run_search, run_extract and run_email_validation. Tool names are the REST endpoint name with underscores.

Run one cheap lookup. The cheapest call in the catalog is email validation at 1 point. It is the one used in the next section, and a single check costs a fraction of a cent.

If Claude cannot see the tools, or a call fails, the troubleshooting table below covers the usual causes.

REST and MCP: one real tool call

Here the MCP side leads. The source is email validation, which checks whether an address is deliverable and returns the verdict, whether the domain accepts all mail (a catch-all), the mail provider and a confidence rating.

You do not write the tool call yourself. You type a question:

Is sam@openai.com a deliverable address? Use NeuralVerge.

Claude decides this needs the email validation tool and sends arguments like these:

{
  "tool": "run_email_validation",
  "arguments": { "email": "sam@openai.com" }
}

The same lookup over REST is one request, for code that already knows when to call:

curl -X POST https://api.neuralverge.ai/functions/v1/run-email-validation \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "email": "sam@openai.com" }'

What comes back is the captured sample response from the source page, copied unchanged:

{
  "session_id": "11111111-1111-1111-1111-111111111111",
  "kind": "email_validation",
  "email": "sam@openai.com",
  "human": "# Email validation\n\n- **Email:** sam@openai.com\n- **Result:** valid\n- **Message:** Mailbox exists and accepts mail.\n- **Catch-all:** No\n- **Provider:** google_workspace\n- **Confidence:** high",
  "machine": {
    "email": "sam@openai.com",
    "result": "valid",
    "message": "Mailbox exists and accepts mail.",
    "isCatchAll": false,
    "provider": "google_workspace",
    "confidence": "high"
  },
  "total_points": 1
}

Claude reads that and answers in a sentence or two. How the tool call and result are displayed depends on the Claude app you use, so we are not describing screens we have not documented.

How to read the response

• —human is a Markdown summary meant to be read. machine is the same result as typed fields for code or an agent. Both are in every response.
• —result, message and confidence are the verdict, an explanation and how strongly it is supported.
• —isCatchAll matters more than it looks. On a catch-all domain a positive result proves less, because the domain accepts everything. Ask Claude to report that field with the verdict.
• —total_points is what the call cost: 1.
• —The session_id in the published sample is a placeholder value.

REST or MCP

The response shape is identical over both. With REST you write the call; with MCP you ask a question in plain English and the model picks the tool. If a failure happens, it comes back inside the tool result flagged as an error (bad parameters, a usage limit, an upstream error), so Claude can read it and react instead of the chat breaking. AI research and saved AI Agents are REST-only.

A small illustrative case

Acme Oy, a Finnish company, appears in a sales lead. You ask Claude to check the contact address before anyone writes to it. Claude calls email validation once, gets the verdict and the catch-all flag, and tells you whether to trust the address or treat it as unverified. The company and address are illustrative; the response shape is the one above.

Cost and limits

Pricing is per resolved lookup, and the interface does not change it. NeuralVerge's documentation is explicit that choosing MCP over REST is an integration decision, not a cost decision. For the example above:

Other sources cost more. Some are 1 point, some are 5, and a few are priced differently, so check the source page before asking Claude to run something in a loop. The pricing page lists action costs, and each source page states its own points and billing line. The email validation page does not state a rule for empty results, so we do not claim one; other source pages say when a lookup that finds nothing is free.

Cost when Claude chooses the calls

A broad question can make Claude run several lookups, and each resolved lookup bills. Start with narrow questions until you know how your prompts translate into calls. A 402 Payment Required means your organization reached its plan or usage limit; Claude sees that inside the tool result and can tell you instead of retrying blindly.

AI research is not an MCP tool

A research run is asynchronous, and an MCP tool has to return inside the call. So AI research and saved AI Agents are started over the REST API; Claude can work with the output afterwards. Search, AI extraction and the data sources are the MCP tools.

Tool count and context

The server lists every tool in a single response, about 150 KB of definitions. Some clients cap how many tools they load, so you may see only part of the catalog. Enable the tools you use or pick a client that lets you choose. A long tool list also takes room in the model's context.

The model chooses the call

That is the point of MCP and also the risk. A vague prompt can produce a wrong or unnecessary call. Name the source and give the identifier when you can: "validate this email with NeuralVerge" beats "check this person out".

Revoking is separate from removing

Removing the connector in Claude does not revoke its access. Every signed-in client appears under Settings → API in the NeuralVerge dashboard with a label starting with Connector:; revoke it there to cut it off immediately.

Troubleshooting

How to decide

Go through this before you hand the connector to a team.

1. —Is sign-in set up for the right account? The connection bills to the organization you signed in with.
2. —Does your client show the tools you expect? Ask it to list them and compare against the source pages you plan to use.
3. —What does each source cost? A 1-point and a 5-point source behave very differently in a loop.
4. —Do you know how to revoke it? Find the Connector entry under Settings → API once, so you are not hunting for it during an incident.
5. —Is research on the REST side? If your plan includes AI research, plan a REST path for it.
6. —Is the personal data handled properly? Email addresses and profiles are personal data. Use them under your own lawful basis and honour opt-outs.

The workflows where a connector fits best are conversations rather than scripts:

• —Account research before a call: pull company records, summarize them, and keep talking about the account in the same chat.
• —Checking lists in place: paste a short list of addresses and have Claude validate them, reporting catch-all cases separately.
• —Registry lookups in plain language: ask for the registered address of a named company in a country's registry.
• —Prototyping an agent: test which sources it should use, with real responses, before writing the REST calls for production. Because both interfaces return the same shape, a call that works in a chat is a good starting point for the equivalent request in code.

Frequently asked questions

Do I need to build or host an MCP server to use NeuralVerge with Claude?

No. NeuralVerge hosts the server at https://mcp.neuralverge.ai/mcp. You add that URL to Claude as a custom connector, sign in, and the tools are available. There is nothing to deploy and no key to paste into a config file.

Do I need a NeuralVerge account to connect?

Yes. Calling a tool needs a sign-in, and every connection bills to your organization. Listing the tools works without credentials, so a client can show the catalog first, but a call without a sign-in is answered with a 401 that points the client to sign-in.

Can Claude run AI research through the MCP server?

No. A research run is asynchronous and an MCP tool has to return inside the call, so AI research and saved AI Agents are started over the REST API. Search, AI extraction and the data sources are the MCP tools.

Does it cost more to call a source from Claude than over REST?

No. Pricing is per resolved lookup regardless of which interface triggered the call, so choosing MCP over REST is an integration decision, not a cost decision. The email validation example in this tutorial costs 1 point, which the source page lists as $1 per 1,000.

Why does Claude show fewer tools than the catalog has?

The server lists every tool in one response, about 150 KB of definitions, and some clients cap how many tools they load. If only some tools are visible, enable the ones you need or use a client that lets you pick tools.

How do I disconnect NeuralVerge from Claude?

Every client that signed in appears under Settings, then API, in the NeuralVerge dashboard with a label starting with Connector. Revoke it there to cut that client off immediately. Removing the connector from Claude alone does not revoke its access.