StackScan MCP server

Ask your assistant what a website runs, and who is behind it.

A Model Context Protocol server for the StackScan Tech Lookup API. Point Claude, Cursor, or any other MCP client at it and you can ask things like:

What is example.com built on?

Who's behind vercel.com? Where are they, what industry, what's their LinkedIn?

How many sites run Klaviyo, and in which countries?

Read-only by design. Nothing here writes to any system of yours.

Install

You need a StackScan account. Sign in, open API Tokens in the dashboard, create a token, and copy your workspace UUID from the same page.

Claude Desktop

Settings → Developer → Edit Config, then add:

{
  "mcpServers": {
    "stackscan": {
      "command": "npx",
      "args": ["-y", "@stackscan/mcp-server"],
      "env": {
        "STACKSCAN_API_TOKEN": "your-token",
        "STACKSCAN_TENANT_ID": "your-workspace-uuid"
      }
    }
  }
}

Restart Claude Desktop. The tools appear under the connectors icon.

Claude Code

claude mcp add stackscan \
  --env STACKSCAN_API_TOKEN=your-token \
  --env STACKSCAN_TENANT_ID=your-workspace-uuid \
  -- npx -y @stackscan/mcp-server

Cursor, Windsurf, and others

Same shape: command: npx, args: ["-y", "@stackscan/mcp-server"], and the two environment variables. Check your client's own MCP docs for where its config file lives.

Tools

A lookup that finds nothing is not charged. Neither is check_credits.

Breadth vs depth

There are two technology tools and they answer different questions. lookup_domains_technologies is the breadth view: many domains, technology names only, ideal for "which of these run Shopify?". lookup_domain_technologies is the depth view: one domain, every technology with its category and how many sites use it globally. The batch tool reports how many technologies it left out per domain, so the model can offer to drill in.

Why the batch tools stop at 20

The REST endpoints behind them take 100 domains per request, and these tools deliberately do not. A tool result goes straight into the model's context, and a hundred full company payloads is tens of thousands of tokens, which crowds out the conversation you are actually having, and the model then has to re-read all of it to answer anything. Twenty compact rows is a table a model can reason over.

It also refuses rather than truncates. If a batch would cost more than your session cap allows, it charges nothing and tells you how many you can afford. Quietly dropping domains would hand back an answer that looks complete and is not.

If you genuinely need hundreds of domains, that is what the REST endpoint is for.

Configuration

The session cap is a spend cap, and it is deliberate

StackScan's own limit is a rate limit: requests per minute. That stops a burst. It does nothing to stop a patient agent quietly spending an entire credit balance over an afternoon, which is a failure mode humans clicking buttons do not have.

So this server keeps its own counter: 25 credit-consuming lookups per session by default, after which it refuses and says why. Misses and check_credits never count against it, because they are never charged. Every successful result carries a footer showing lookups remaining and your last known balance, so the model can pace itself instead of discovering the limit by hitting it.

If you mean to go higher, raise it deliberately ("STACKSCAN_SESSION_LOOKUP_CAP": "200") rather than finding out afterwards.

Development

npm install
npm run build

test-stdio.mjs is a minimal JSON-RPC client that drives the server over a real stdio pipe, so you can exercise it without an MCP host:

export STACKSCAN_API_TOKEN=...
export STACKSCAN_TENANT_ID=...
node test-stdio.mjs check_credits lookup_company:stripe.com

Never write to stdout in this server. stdout is the JSON-RPC channel; anything else on it corrupts the protocol. Diagnostics go to stderr.

Notes for anyone reading the source

Three things about the upstream API shaped this code and are easy to get wrong:

A miss is HTTP 200, not 404. The API answers "no data for that domain" with 200 {"success": false, "error": "No data available"}. A client that trusts the status code reads the absent payload and crashes. Every response goes through one place that separates a miss from a failure, so only a real failure is reported as an error.

Cached responses are still charged. The API's response cache shields its database, not your wallet: a repeat lookup inside the cache window costs the same credit. That is why the local balance decrements on every hit instead of trying to guess which calls were free.

Two base URLs are live. The API's home is api.stackscan.com/v1/.... The older app.stackscan.com/api/v1/... paths are still served and are deliberately not redirected, because a redirect would drop the body on POST. If your config still names the old host, this server detects it and keeps using the old path shape, so upgrading will not break you.

Links

• API reference
• What this is, in plain English
• StackScan

License

MIT. See LICENSE.

The licence covers this client only, roughly 500 lines that make HTTP requests. The StackScan data it reaches is a paid service and is not covered by it.