Neuralverge MCP Server

NeuralVerge MCP Server

The data layer for agentic AI.

One API and MCP server that gives an agent verified, structured, cited facts about the real world, mid-task. Agents don't need search results or raw HTML — they need a finished answer, from one API instead of five vendors.

NeuralVerge draws structured data from 300+ sources — e-commerce marketplaces, corporate registries, startup and funding data, company intelligence, product reviews, enrichment and social — plus the open web, all in one request format. This MCP (Model Context Protocol) server exposes Search, AI Extract and every Data Source as tools, so any MCP-compatible client (Claude, ChatGPT, Gemini, Cursor and others) can call them directly. No scrapers, no pipelines, no glue code. Plans from $50/month.

Why NeuralVerge

• The real world, structured. E-commerce, company and people data from marketplaces, official registries and vertical sources, not just the web — all returned as JSON.
• One API instead of five. Search, extraction and enrichment on one key and one balance, with one request format and no per-source fees or separate contracts.
• Agent-native. Tools are available to the agent mid-task over MCP, with no glue code.
• Cost you control. Points-based pricing: extraction and search at 5 points per call, plans from $50/month (see Pricing).

Teams use NeuralVerge for agentic products, go-to-market, outbound, market research, competitor analysis, brand monitoring, public procurement, e-commerce intelligence, due diligence, KYC/KYB and recruiting.

Getting an API key

Before installing, you need a NeuralVerge account and an API key:

1. Sign up at app.neuralverge.ai (or via neuralverge.ai → "Get started").
2. Choose a plan (see Pricing below) — this determines your monthly points allowance.
3. Generate an API key from your account/API settings in the app.
4. Use that key as NEURALVERGE_API_KEY in the Configuration step further down.

MCP server overview

NeuralVerge exposes its data layer as a remote Model Context Protocol (MCP) server, so any MCP-compatible client — Claude, ChatGPT, Gemini, Cursor, or a custom agent — can call Search, AI Extract, and every Data Source (including marketplaces and product reviews) directly as tools, without writing any HTTP integration code. See the official page: docs.neuralverge.ai/mcp-server.

This repository is a self-hosted equivalent of that server: same tool names, same input schemas, same behavior, and it can run either as a local stdio process or as a Streamable HTTP server. Notes:

• AI Research is not available over MCP: a research run is asynchronous, while an MCP tool has to return its result inside the call. Start research runs over the REST API instead.
• Total tool count: 342 — run_search, run_extract and every Data Source endpoint.

Quickstart

1. Get an API key. In the NeuralVerge dashboard open Settings → API, click Create API Key, and copy the key. Keep it private — it is the only thing that authenticates the connection.

2. Point your client at a server. Either use NeuralVerge's hosted endpoint, or run this repository (see Installation and Transports & client compatibility).

   Hosted endpoint, Claude Code (sign in with OAuth from /mcp; add --header "Authorization: Bearer YOUR_API_KEY" for CI or headless use):

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

   Hosted endpoint, Cursor and other JSON clients (omit headers to sign in with OAuth instead of an API key):

   {
     "mcpServers": {
       "neuralverge": {
         "url": "https://mcp.neuralverge.ai/mcp",
         "headers": {
           "Authorization": "Bearer YOUR_API_KEY"
         }
       }
     }
   }
   

   This repository, Claude Code (local stdio):

   claude mcp add neuralverge -e NEURALVERGE_API_KEY=YOUR_API_KEY -- node /absolute/path/to/MCP/dist/index.js
   

3. Let the agent call tools. Start a new conversation and ask for something that needs data, for example: "Find the registered address of Škoda Auto in the Czech company registry." The agent picks the right tool, calls it mid-task, and answers from the result.

Hosted endpoint

https://mcp.neuralverge.ai/mcp

The hosted server speaks MCP's Streamable HTTP transport and is stateless — each request is handled independently, with no session to keep alive between calls. Requests must include:

Accept: application/json, text/event-stream

The earlier endpoint, https://api.neuralverge.ai/functions/v1/mcp-server, still works with an API key but does not support sign-in. Use https://mcp.neuralverge.ai/mcp for new connections.

Authentication

The hosted server accepts two kinds of credentials, both sent as Authorization: Bearer … and both billed to your organization:

• Sign in with OAuth (recommended). Clients that support MCP authorization — Claude, ChatGPT, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI and others — discover NeuralVerge's sign-in automatically: add the URL with no credentials and approve access in the browser window that opens (OAuth 2.1, dynamic client registration, PKCE S256). Each connection appears under Settings → API in the dashboard with a label starting with Connector:; the key does not expire, so revoke it there to disconnect a client. Removing the server from the client alone does not revoke it.
• API key. For clients without browser sign-in, create a key under Settings → API (see docs.neuralverge.ai/authentication) and send Authorization: Bearer YOUR_API_KEY — the same key the REST API uses.

Listing tools (initialize, tools/list) works without credentials; calling a tool requires them, otherwise the hosted server answers 401 Unauthorized with a WWW-Authenticate header pointing OAuth-capable clients to the sign-in flow. When you run this repository instead, the key is passed to the server process as NEURALVERGE_API_KEY (see Configuration); there is no OAuth flow.

Tool count

tools/list returns all 342 tools in one response — several hundred kilobytes of JSON. Three things follow from that:

• A model pays for those definitions in context on every conversation that has the server attached.
• Clients that limit the number of active tools will load only some of them, and the rest will be invisible to the model. Windsurf, for instance, allows 100 at a time.
• Every tool is stateless, read-only and independent of the others (and is annotated readOnlyHint here), so a client can run several calls in parallel.

If you only need a few sources, attach the server to a client that lets you pick tools, or proxy it and expose the subset you use.

Synchronous results

Every tool returns its result directly in the tool response — there is no session to poll. A scrape that hits its time budget returns partial: true with the records collected so far, so a long run degrades to a shorter result instead of failing.

Connecting an MCP client

• Claude Code, Cursor and other JSON-config clients — see the Quickstart above.
• Claude, ChatGPT — add a custom connector / MCP server pointing at the hosted endpoint above and choose OAuth (leave client ID and secret empty; the client registers itself). To use your own deployment of this server in HTTP mode instead, see Streamable HTTP.
• Any other Streamable HTTP client — point it at the URL; if it supports MCP authorization it will offer to sign in, otherwise set Authorization: Bearer YOUR_API_KEY. Clients that only speak SSE or stdio can use an MCP proxy such as mcp-remote. Step-by-step setup for every client is on Connect MCP.
• MCP Inspector — run npx @modelcontextprotocol/inspector, choose the Streamable HTTP transport, connect to https://mcp.neuralverge.ai/mcp, and sign in with OAuth or set the Authorization header to explore and test tools interactively.

Example request

curl -X POST "https://mcp.neuralverge.ai/mcp" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "run_email_validation",
      "arguments": { "email": "john@example.com" }
    }
  }'

Errors

A tool call without an Authorization header fails the whole request with 401 Unauthorized; a key that is invalid or revoked is reported inside the tool result instead — remove and re-add the server in your client to sign in again. Once a request is authenticated, errors from an individual tool call (invalid parameters, 402 Payment Required for usage limits, upstream failures, and so on) are returned inside the MCP tool result with isError: true rather than as an HTTP failure — this lets the calling model see and react to the error instead of the whole request failing. The underlying status codes are the same ones described in Errors & Rate Limits and in Error handling below.

Tools

The server exposes 342 tools — one per NeuralVerge API endpoint (Search, AI Extract and every Data Source), as listed below. AI Research is not available over MCP. The tool list is generated from the official OpenAPI spec, so input schemas (types, enums, min/max, required fields) match the REST API exactly.

Every tool returns its result synchronously (a scrape that hits its time budget returns partial: true with the records collected so far). Tool names follow the endpoint path: POST /functions/v1/run-amazon-product-search → run_amazon_product_search.

Context size: all 342 tool definitions together are large. If your client caps active tools or you want to save context, enable only the tools you need in the client's tool picker.

All available tools

342 tools, generated by npm run generate from the NeuralVerge OpenAPI spec. Required parameters are marked with *.

Search (1)

AI Extract (1)

Social media (5)

Enrichment (5)

Company intelligence (1)

E-commerce & Marketplaces (128)

Documentation truncated — see the full README on GitHub.